工作流 /
图像 /
Minimax H3 Image Studio 8k+ (12+ GB VRAM)

Minimax H3 Image Studio 8k+ (12+ GB VRAM)
Minimax H3 Image Studio 8k+ (12+ GB VRAM)
V2.0 (Adaptation for official H3 Image)MiniMax H3 Image Studio – 8K+ (Ref2VA + Turbo Mode + SeedVR2)Text-to-image / reference-based image generation with MiniMax H3 image. Includes an optional SeedVR2 4K+ refine pass for…
图像
42 个节点 6 次使用 Hopik65 118 KB
需要的权重
这个工作流引用了 1 个权重文件,其中 0 个能在本站模型库里找到。
Minimax_H3_Image_Studio_8k.json
模型库暂未收录
节点构成
LoadImage
× 9
PrimitiveStringMultiline
× 3
Note
× 3
CLIPLoader
× 2
SaveImage
× 2
UNETLoader
× 1
VAELoader
× 1
H3ImageResolutionPreset
× 1
SeedVR2LoadVAEModel
× 1
SeedVR2LoadDiTModel
× 1
SeedVR2VideoUpscaler
× 1
Image Comparer (rgthree)
× 1
Fast Groups Muter (rgthree)
× 1
MarkdownNote
× 1
KSampler
× 1
VAEDecode
× 1
官方标注的模型
ema_vae_fp16.safetensorsminimax_h3_image_pruned_int8_convrot.safetensorsminimax_h3_image_vae_fp16.safetensorsqwen3.5_9b_w4a8_gptq.safetensorsqwen3vl_32b_minimax_h3_nvfp4_awq.safetensorsseedvr2_ema_7b-Q4_K_M.gguf
工作流 JSON
和上面「下载工作流 JSON」是同一份文件。
1{
2 "id": "5b8de34a-144a-4de1-a05a-ec2b57bf45f6",
3 "revision": 0,
4 "last_node_id": 75,
5 "last_link_id": 171,
6 "nodes": [
7 {
8 "id": 1,
9 "type": "UNETLoader",
10 "pos": [
11 -56.904295948872154,
12 360.888695758076
13 ],
14 "size": [
15 390,
16 82
17 ],
18 "flags": [],
19 "order": 0,
20 "mode": 0,
21 "inputs": [],
22 "outputs": [
23 {
24 "name": "MODEL",
25 "type": "MODEL",
26 "links": [
27 170
28 ]
29 }
30 ],
31 "properties": {
32 "cnr_id": "comfy-core",
33 "ver": "0.34.0",
34 "Node name for S&R": "UNETLoader"
35 },
36 "widgets_values": [
37 "minimax_h3_image_pruned_int8_convrot.safetensors",
38 "default"
39 ],
40 "widgets_values_named": {
41 "unet_name": "minimax_h3_image_pruned_int8_convrot.safetensors",
42 "weight_dtype": "default"
43 },
44 "color": "#322",
45 "bgcolor": "#533"
46 },
47 {
48 "id": 2,
49 "type": "CLIPLoader",
50 "pos": [
51 -55.42047183254877,
52 495.8172506918854
53 ],
54 "size": [
55 378.15960008918705,
56 106.74002499442577
57 ],
58 "flags": [],
59 "order": 1,
60 "mode": 0,
61 "inputs": [],
62 "outputs": [
63 {
64 "name": "CLIP",
65 "type": "CLIP",
66 "links": [
67 171
68 ]
69 }
70 ],
71 "properties": {
72 "cnr_id": "comfy-core",
73 "ver": "0.34.0",
74 "Node name for S&R": "CLIPLoader"
75 },
76 "widgets_values": [
77 "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
78 "minimax",
79 "default"
80 ],
81 "widgets_values_named": {
82 "clip_name": "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
83 "type": "minimax",
84 "device": "default"
85 },
86 "color": "#322",
87 "bgcolor": "#533"
88 },
89 {
90 "id": 3,
91 "type": "VAELoader",
92 "pos": [
93 -40.24995944681716,
94 681.2260293563323
95 ],
96 "size": [
97 390,
98 58
99 ],
100 "flags": [],
101 "order": 2,
102 "mode": 0,
103 "inputs": [],
104 "outputs": [
105 {
106 "name": "VAE",
107 "type": "VAE",
108 "links": [
109 90
110 ]
111 }
112 ],
113 "properties": {
114 "cnr_id": "comfy-core",
115 "ver": "0.34.0",
116 "Node name for S&R": "VAELoader"
117 },
118 "widgets_values": [
119 "minimax_h3_image_vae_fp16.safetensors"
120 ],
121 "widgets_values_named": {
122 "vae_name": "minimax_h3_image_vae_fp16.safetensors"
123 },
124 "color": "#322",
125 "bgcolor": "#533"
126 },
127 {
128 "id": 13,
129 "type": "SaveImage",
130 "pos": [
131 2154.0021873485775,
132 452.2096700940575
133 ],
134 "size": [
135 315.27303310845036,
136 282
137 ],
138 "flags": [],
139 "order": 34,
140 "mode": 0,
141 "inputs": [
142 {
143 "name": "images",
144 "type": "IMAGE",
145 "link": 100
146 }
147 ],
148 "outputs": [
149 {
150 "name": "images",
151 "type": "IMAGE",
152 "links": [
153 153
154 ]
155 }
156 ],
157 "properties": {
158 "cnr_id": "comfy-core",
159 "ver": "0.34.0",
160 "Node name for S&R": "SaveImage"
161 },
162 "widgets_values": [
163 "MiniMax_H3_Image"
164 ],
165 "widgets_values_named": {
166 "filename_prefix": "MiniMax_H3_Image"
167 }
168 },
169 {
170 "id": 4,
171 "type": "H3ImageResolutionPreset",
172 "pos": [
173 815.385676809357,
174 377.8542647570573
175 ],
176 "size": [
177 390,
178 170
179 ],
180 "flags": {
181 "collapsed": false
182 },
183 "order": 3,
184 "mode": 0,
185 "inputs": [
186 {
187 "name": "source_image",
188 "shape": 7,
189 "type": "IMAGE",
190 "link": null
191 }
192 ],
193 "outputs": [
194 {
195 "name": "width",
196 "type": "INT",
197 "links": [
198 92
199 ]
200 },
201 {
202 "name": "height",
203 "type": "INT",
204 "links": [
205 93
206 ]
207 },
208 {
209 "name": "resolution_info",
210 "type": "STRING",
211 "links": null
212 }
213 ],
214 "properties": {
215 "Image Studio": "v18",
216 "cnr_id": "minimax-h3-image-studio",
217 "ver": "b4a233911324f3e00dbd13aa4f65475e78c482d8",
218 "Node name for S&R": "H3ImageResolutionPreset",
219 "aux_id": "astropuzzo/ComfyUI-MiniMax-H3-Image-Studio"
220 },
221 "widgets_values": [
222 "4:3 landscape",
223 "custom megapixels",
224 2,
225 false
226 ],
227 "widgets_values_named": {
228 "aspect_ratio": "4:3 landscape",
229 "resolution_profile": "custom megapixels",
230 "custom_megapixels": 2,
231 "limit_to_native_area": false
232 },
233 "color": "#443066",
234 "bgcolor": "#211635"
235 },
236 {
237 "id": 0,
238 "type": "LoadImage",
239 "pos": [
240 -48.99631368564453,
241 823.3924540067654
242 ],
243 "size": [
244 390,
245 314.0000000000001
246 ],
247 "flags": {
248 "collapsed": false
249 },
250 "order": 4,
251 "mode": 2,
252 "inputs": [],
253 "outputs": [
254 {
255 "name": "IMAGE",
256 "type": "IMAGE",
257 "links": [
258 101,
259 127
260 ]
261 },
262 {
263 "name": "MASK",
264 "type": "MASK",
265 "links": null
266 }
267 ],
268 "title": "Image 1",
269 "properties": {
270 "cnr_id": "comfy-core",
271 "ver": "0.34.0",
272 "Node name for S&R": "LoadImage"
273 },
274 "widgets_values": [
275 "{6B6C06F8-BDD8-4EE4-B6D6-CBEFD0FB4490}.png",
276 "image"
277 ],
278 "widgets_values_named": {
279 "image": "{6B6C06F8-BDD8-4EE4-B6D6-CBEFD0FB4490}.png",
280 "upload": "image"
281 }
282 },
283 {
284 "id": 14,
285 "type": "LoadImage",
286 "pos": [
287 -40.12579555501318,
288 1220.032993585685
289 ],
290 "size": [
291 381.04569756744763,
292 272.48459781271185
293 ],
294 "flags": {
295 "collapsed": false
296 },
297 "order": 5,
298 "mode": 2,
299 "inputs": [],
300 "outputs": [
301 {
302 "name": "IMAGE",
303 "type": "IMAGE",
304 "links": [
305 102,
306 128
307 ]
308 },
309 {
310 "name": "MASK",
311 "type": "MASK",
312 "links": null
313 }
314 ],
315 "title": "Image 2",
316 "properties": {
317 "cnr_id": "comfy-core",
318 "ver": "0.34.0",
319 "Node name for S&R": "LoadImage"
320 },
321 "widgets_values": [
322 "images.jpg",
323 "image"
324 ],
325 "widgets_values_named": {
326 "image": "images.jpg",
327 "upload": "image"
328 }
329 },
330 {
331 "id": 29,
332 "type": "SeedVR2LoadVAEModel",
333 "pos": [
334 2528.1460578218125,
335 494.89414444358914
336 ],
337 "size": [
338 317.9141446387274,
339 317.46834137437736
340 ],
341 "flags": [],
342 "order": 6,
343 "mode": 2,
344 "inputs": [
345 {
346 "name": "torch_compile_args",
347 "shape": 7,
348 "type": "TORCH_COMPILE_ARGS",
349 "link": null
350 }
351 ],
352 "outputs": [
353 {
354 "name": "SEEDVR2_VAE",
355 "type": "SEEDVR2_VAE",
356 "links": [
357 61
358 ]
359 }
360 ],
361 "properties": {
362 "cnr_id": "seedvr2_videoupscaler",
363 "ver": "690cc39379c1481159ddd451368dbf2295930fc6",
364 "Node name for S&R": "SeedVR2LoadVAEModel",
365 "aux_id": "ainvfx/ComfyUI-SeedVR2_VideoUpscaler",
366 "ue_properties": {
367 "version": "7.0.1",
368 "widget_ue_connectable": {
369 "model": true,
370 "device": true,
371 "encode_tiled": true,
372 "encode_tile_size": true,
373 "encode_tile_overlap": true,
374 "decode_tiled": true,
375 "decode_tile_size": true,
376 "decode_tile_overlap": true,
377 "tile_debug": true,
378 "offload_device": true,
379 "cache_model": true
380 }
381 }
382 },
383 "widgets_values": [
384 "ema_vae_fp16.safetensors",
385 "cuda:0",
386 true,
387 256,
388 32,
389 true,
390 256,
391 32,
392 "false",
393 "cpu",
394 false
395 ],
396 "widgets_values_named": {
397 "model": "ema_vae_fp16.safetensors",
398 "device": "cuda:0",
399 "encode_tiled": true,
400 "encode_tile_size": 256,
401 "encode_tile_overlap": 32,
402 "decode_tiled": true,
403 "decode_tile_size": 256,
404 "decode_tile_overlap": 32,
405 "tile_debug": "false",
406 "offload_device": "cpu",
407 "cache_model": false
408 }
409 },
410 {
411 "id": 30,
412 "type": "SeedVR2LoadDiTModel",
413 "pos": [
414 2868.528443513793,
415 360.00174588541984
416 ],
417 "size": [
418 307.6646484375,
419 202
420 ],
421 "flags": [],
422 "order": 7,
423 "mode": 2,
424 "inputs": [
425 {
426 "name": "torch_compile_args",
427 "shape": 7,
428 "type": "TORCH_COMPILE_ARGS",
429 "link": null
430 }
431 ],
432 "outputs": [
433 {
434 "name": "SEEDVR2_DIT",
435 "type": "SEEDVR2_DIT",
436 "links": [
437 60
438 ]
439 }
440 ],
441 "properties": {
442 "cnr_id": "seedvr2_videoupscaler",
443 "ver": "690cc39379c1481159ddd451368dbf2295930fc6",
444 "Node name for S&R": "SeedVR2LoadDiTModel",
445 "aux_id": "ainvfx/ComfyUI-SeedVR2_VideoUpscaler",
446 "ue_properties": {
447 "version": "7.0.1",
448 "widget_ue_connectable": {
449 "model": true,
450 "device": true,
451 "blocks_to_swap": true,
452 "swap_io_components": true,
453 "offload_device": true,
454 "cache_model": true,
455 "attention_mode": true
456 }
457 }
458 },
459 "widgets_values": [
460 "seedvr2_ema_7b-Q4_K_M.gguf",
461 "cuda:0",
462 36,
463 true,
464 "cpu",
465 false,
466 "sdpa"
467 ],
468 "widgets_values_named": {
469 "model": "seedvr2_ema_7b-Q4_K_M.gguf",
470 "device": "cuda:0",
471 "blocks_to_swap": 36,
472 "swap_io_components": true,
473 "offload_device": "cpu",
474 "cache_model": false,
475 "attention_mode": "sdpa"
476 }
477 },
478 {
479 "id": 31,
480 "type": "SeedVR2VideoUpscaler",
481 "pos": [
482 2879.1943298605224,
483 612.5343775000969
484 ],
485 "size": [
486 289.06908246036255,
487 386
488 ],
489 "flags": [],
490 "order": 38,
491 "mode": 2,
492 "inputs": [
493 {
494 "name": "image",
495 "type": "IMAGE",
496 "link": 155
497 },
498 {
499 "name": "dit",
500 "type": "SEEDVR2_DIT",
501 "link": 60
502 },
503 {
504 "name": "vae",
505 "type": "SEEDVR2_VAE",
506 "link": 61
507 },
508 {
509 "name": "resolution",
510 "type": "INT",
511 "widget": {
512 "name": "resolution"
513 },
514 "link": 169
515 }
516 ],
517 "outputs": [
518 {
519 "name": "IMAGE",
520 "type": "IMAGE",
521 "links": [
522 83,
523 157
524 ]
525 }
526 ],
527 "properties": {
528 "cnr_id": "seedvr2_videoupscaler",
529 "ver": "690cc39379c1481159ddd451368dbf2295930fc6",
530 "Node name for S&R": "SeedVR2VideoUpscaler",
531 "aux_id": "ainvfx/ComfyUI-SeedVR2_VideoUpscaler",
532 "ue_properties": {
533 "version": "7.0.1",
534 "widget_ue_connectable": {
535 "seed": true,
536 "resolution": true,
537 "max_resolution": true,
538 "batch_size": true,
539 "uniform_batch_size": true,
540 "color_correction": true,
541 "temporal_overlap": true,
542 "prepend_frames": true,
543 "input_noise_scale": true,
544 "latent_noise_scale": true,
545 "offload_device": true,
546 "enable_debug": true
547 }
548 }
549 },
550 "widgets_values": [
551 1102,
552 "fixed",
553 4096,
554 0,
555 1,
556 false,
557 "lab",
558 0,
559 0,
560 0,
561 0,
562 "cpu",
563 false
564 ],
565 "widgets_values_named": {
566 "seed": 1102,
567 "control_after_generate": "fixed",
568 "resolution": 4096,
569 "max_resolution": 0,
570 "batch_size": 1,
571 "uniform_batch_size": false,
572 "color_correction": "lab",
573 "temporal_overlap": 0,
574 "prepend_frames": 0,
575 "input_noise_scale": 0,
576 "latent_noise_scale": 0,
577 "offload_device": "cpu",
578 "enable_debug": false
579 }
580 },
581 {
582 "id": 32,
583 "type": "Image Comparer (rgthree)",
584 "pos": [
585 3195.0321572400317,
586 523.2393919072591
587 ],
588 "size": [
589 513.1261135479682,
590 478.9293805454947
591 ],
592 "flags": [],
593 "order": 41,
594 "mode": 0,
595 "inputs": [
596 {
597 "dir": 3,
598 "name": "image_a",
599 "type": "IMAGE",
600 "link": 156
601 },
602 {
603 "dir": 3,
604 "name": "image_b",
605 "type": "IMAGE",
606 "link": 158
607 }
608 ],
609 "outputs": [
610 {
611 "dir": 4,
612 "name": "images",
613 "shape": 3,
614 "type": "IMAGE",
615 "links": []
616 }
617 ],
618 "title": "Image Comparer (rgthree) 7BQ8",
619 "properties": {
620 "cnr_id": "rgthree-comfy",
621 "ver": "1.0.2508012353",
622 "comparer_mode": "Slide",
623 "ue_properties": {
624 "version": "7.0.1",
625 "widget_ue_connectable": []
626 }
627 },
628 "widgets_values": [
629 [
630 {
631 "name": "A",
632 "selected": true,
633 "url": "/api/view?filename=rgthree.compare._temp_nxarc_00005_.png&type=temp&subfolder=&rand=0.9991026376243927"
634 },
635 {
636 "name": "B",
637 "selected": true,
638 "url": "/api/view?filename=rgthree.compare._temp_nxarc_00006_.png&type=temp&subfolder=&rand=0.6970344726588816"
639 }
640 ]
641 ],
642 "widgets_values_named": {
643 "rgthree_comparer": {
644 "images": [
645 {
646 "name": "A",
647 "selected": true,
648 "url": "/api/view?filename=rgthree.compare._temp_nxarc_00005_.png&type=temp&subfolder=&rand=0.9991026376243927",
649 "img": []
650 },
651 {
652 "name": "B",
653 "selected": true,
654 "url": "/api/view?filename=rgthree.compare._temp_nxarc_00006_.png&type=temp&subfolder=&rand=0.6970344726588816",
655 "img": []
656 }
657 ]
658 }
659 }
660 },
661 {
662 "id": 34,
663 "type": "LoadImage",
664 "pos": [
665 -50.94572574862306,
666 2341.2521717648697
667 ],
668 "size": [
669 390,
670 314
671 ],
672 "flags": {
673 "collapsed": false
674 },
675 "order": 8,
676 "mode": 2,
677 "inputs": [],
678 "outputs": [
679 {
680 "name": "IMAGE",
681 "type": "IMAGE",
682 "links": [
683 105,
684 131
685 ]
686 },
687 {
688 "name": "MASK",
689 "type": "MASK",
690 "links": []
691 }
692 ],
693 "title": "Image 5",
694 "properties": {
695 "cnr_id": "comfy-core",
696 "ver": "0.34.0",
697 "Node name for S&R": "LoadImage"
698 },
699 "widgets_values": [
700 "images.jpg",
701 "image"
702 ],
703 "widgets_values_named": {
704 "image": "images.jpg",
705 "upload": "image"
706 }
707 },
708 {
709 "id": 35,
710 "type": "LoadImage",
711 "pos": [
712 -47.19353927811031,
713 1567.276163971839
714 ],
715 "size": [
716 390,
717 314
718 ],
719 "flags": {
720 "collapsed": false
721 },
722 "order": 9,
723 "mode": 2,
724 "inputs": [],
725 "outputs": [
726 {
727 "name": "IMAGE",
728 "type": "IMAGE",
729 "links": [
730 103,
731 129
732 ]
733 },
734 {
735 "name": "MASK",
736 "type": "MASK",
737 "links": []
738 }
739 ],
740 "title": "Image 3",
741 "properties": {
742 "cnr_id": "comfy-core",
743 "ver": "0.34.0",
744 "Node name for S&R": "LoadImage"
745 },
746 "widgets_values": [
747 "images.jpg",
748 "image"
749 ],
750 "widgets_values_named": {
751 "image": "images.jpg",
752 "upload": "image"
753 }
754 },
755 {
756 "id": 36,
757 "type": "LoadImage",
758 "pos": [
759 -45.20338044027696,
760 1960.4846883931234
761 ],
762 "size": [
763 387.5536363820635,
764 314
765 ],
766 "flags": {
767 "collapsed": false
768 },
769 "order": 10,
770 "mode": 2,
771 "inputs": [],
772 "outputs": [
773 {
774 "name": "IMAGE",
775 "type": "IMAGE",
776 "links": [
777 104,
778 130
779 ]
780 },
781 {
782 "name": "MASK",
783 "type": "MASK",
784 "links": []
785 }
786 ],
787 "title": "Image 4",
788 "properties": {
789 "cnr_id": "comfy-core",
790 "ver": "0.34.0",
791 "Node name for S&R": "LoadImage"
792 },
793 "widgets_values": [
794 "images.jpg",
795 "image"
796 ],
797 "widgets_values_named": {
798 "image": "images.jpg",
799 "upload": "image"
800 }
801 },
802 {
803 "id": 37,
804 "type": "LoadImage",
805 "pos": [
806 350.87608738956527,
807 828.0811610534995
808 ],
809 "size": [
810 390,
811 314
812 ],
813 "flags": {
814 "collapsed": false
815 },
816 "order": 11,
817 "mode": 2,
818 "inputs": [],
819 "outputs": [
820 {
821 "name": "IMAGE",
822 "type": "IMAGE",
823 "links": [
824 106,
825 132
826 ]
827 },
828 {
829 "name": "MASK",
830 "type": "MASK",
831 "links": []
832 }
833 ],
834 "title": "Image 6",
835 "properties": {
836 "cnr_id": "comfy-core",
837 "ver": "0.34.0",
838 "Node name for S&R": "LoadImage"
839 },
840 "widgets_values": [
841 "images.jpg",
842 "image"
843 ],
844 "widgets_values_named": {
845 "image": "images.jpg",
846 "upload": "image"
847 }
848 },
849 {
850 "id": 38,
851 "type": "LoadImage",
852 "pos": [
853 359.7685359160873,
854 1205.0749340183063
855 ],
856 "size": [
857 392.01824998479765,
858 311.3090000202699
859 ],
860 "flags": {
861 "collapsed": false
862 },
863 "order": 12,
864 "mode": 2,
865 "inputs": [],
866 "outputs": [
867 {
868 "name": "IMAGE",
869 "type": "IMAGE",
870 "links": [
871 107,
872 133
873 ]
874 },
875 {
876 "name": "MASK",
877 "type": "MASK",
878 "links": []
879 }
880 ],
881 "title": "Image 7",
882 "properties": {
883 "cnr_id": "comfy-core",
884 "ver": "0.34.0",
885 "Node name for S&R": "LoadImage"
886 },
887 "widgets_values": [
888 "images.jpg",
889 "image"
890 ],
891 "widgets_values_named": {
892 "image": "images.jpg",
893 "upload": "image"
894 }
895 },
896 {
897 "id": 39,
898 "type": "LoadImage",
899 "pos": [
900 361.9014769139223,
901 1569.7421062596138
902 ],
903 "size": [
904 390,
905 314
906 ],
907 "flags": {
908 "collapsed": false
909 },
910 "order": 13,
911 "mode": 2,
912 "inputs": [],
913 "outputs": [
914 {
915 "name": "IMAGE",
916 "type": "IMAGE",
917 "links": [
918 108,
919 134
920 ]
921 },
922 {
923 "name": "MASK",
924 "type": "MASK",
925 "links": []
926 }
927 ],
928 "title": "Image 8",
929 "properties": {
930 "cnr_id": "comfy-core",
931 "ver": "0.34.0",
932 "Node name for S&R": "LoadImage"
933 },
934 "widgets_values": [
935 "images.jpg",
936 "image"
937 ],
938 "widgets_values_named": {
939 "image": "images.jpg",
940 "upload": "image"
941 }
942 },
943 {
944 "id": 40,
945 "type": "LoadImage",
946 "pos": [
947 354.6299140588164,
948 1954.0273752515875
949 ],
950 "size": [
951 390,
952 314
953 ],
954 "flags": {
955 "collapsed": false
956 },
957 "order": 14,
958 "mode": 2,
959 "inputs": [],
960 "outputs": [
961 {
962 "name": "IMAGE",
963 "type": "IMAGE",
964 "links": [
965 109,
966 135
967 ]
968 },
969 {
970 "name": "MASK",
971 "type": "MASK",
972 "links": []
973 }
974 ],
975 "title": "Image 9",
976 "properties": {
977 "cnr_id": "comfy-core",
978 "ver": "0.34.0",
979 "Node name for S&R": "LoadImage"
980 },
981 "widgets_values": [
982 "images.jpg",
983 "image"
984 ],
985 "widgets_values_named": {
986 "image": "images.jpg",
987 "upload": "image"
988 }
989 },
990 {
991 "id": 41,
992 "type": "Fast Groups Muter (rgthree)",
993 "pos": [
994 1279.242965820201,
995 890.3021981018636
996 ],
997 "size": [
998 293.84809166352613,
999 394
1000 ],
1001 "flags": [],
1002 "order": 15,
1003 "mode": 0,
1004 "inputs": [],
1005 "outputs": [
1006 {
1007 "name": "OPT_CONNECTION",
1008 "type": "*",
1009 "links": null
1010 }
1011 ],
1012 "properties": {
1013 "matchColors": "",
1014 "matchTitle": "",
1015 "showNav": true,
1016 "showAllGraphs": true,
1017 "sort": "position",
1018 "customSortAlphabet": "",
1019 "toggleRestriction": "default"
1020 }
1021 },
1022 {
1023 "id": 43,
1024 "type": "SaveImage",
1025 "pos": [
1026 3728.4205098391963,
1027 394.9441399968843
1028 ],
1029 "size": [
1030 390,
1031 270
1032 ],
1033 "flags": [],
1034 "order": 39,
1035 "mode": 2,
1036 "inputs": [
1037 {
1038 "name": "images",
1039 "type": "IMAGE",
1040 "link": 83
1041 }
1042 ],
1043 "outputs": [
1044 {
1045 "name": "images",
1046 "type": "IMAGE",
1047 "links": []
1048 }
1049 ],
1050 "properties": {
1051 "cnr_id": "comfy-core",
1052 "ver": "0.34.0",
1053 "Node name for S&R": "SaveImage"
1054 },
1055 "widgets_values": [
1056 "MiniMax_H3_Image"
1057 ],
1058 "widgets_values_named": {
1059 "filename_prefix": "MiniMax_H3_Image"
1060 }
1061 },
1062 {
1063 "id": 44,
1064 "type": "PrimitiveStringMultiline",
1065 "pos": [
1066 739.7400444609955,
1067 880.388412000533
1068 ],
1069 "size": [
1070 480,
1071 660
1072 ],
1073 "flags": {
1074 "collapsed": false
1075 },
1076 "order": 16,
1077 "mode": 0,
1078 "inputs": [],
1079 "outputs": [
1080 {
1081 "name": "STRING",
1082 "type": "STRING",
1083 "links": [
1084 143
1085 ]
1086 }
1087 ],
1088 "title": "Text (Prompt)",
1089 "properties": {
1090 "cnr_id": "comfy-core",
1091 "ver": "0.44.0",
1092 "Node name for S&R": "PrimitiveStringMultiline"
1093 },
1094 "widgets_values": [
1095 ""
1096 ],
1097 "widgets_values_named": {
1098 "value": ""
1099 },
1100 "color": "#223",
1101 "bgcolor": "#335"
1102 },
1103 {
1104 "id": 45,
1105 "type": "MarkdownNote",
1106 "pos": [
1107 -603.4814273539805,
1108 341.92276428631504
1109 ],
1110 "size": [
1111 520,
1112 770
1113 ],
1114 "flags": [],
1115 "order": 17,
1116 "mode": 0,
1117 "inputs": [],
1118 "outputs": [],
1119 "properties": [],
1120 "widgets_values": [
1121 "## Model Links\n\n- [Hugging Face:Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3)\n- [Hugging Face:Pika-Labs/Pika-H3-Flex-Turbo-LoRA](https://huggingface.co/Pika-Labs/Pika-H3-Flex-Turbo-LoRA)\n\n**diffusion_models**\n\n- [minimax_h3_image_pruned_int8_convrot.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3-Image/resolve/main/diffusion_models/minimax_h3_image_pruned_int8_convrot.safetensors) (21 GB)\n\n**text_encoders**\n\n- [qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors) (14.9 GB)\n\n- [qwen3.5_9b_w4a8_gptq.safetensors](https://huggingface.co/Comfy-Org/Qwen3.5/resolve/main/text_encoders/qwen3.5_9b_w4a8_gptq.safetensors) (6.84 GB)\n\n\n**vae**\n\n- [minimax_h3_image_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3-Image/resolve/main/vae/minimax_h3_image_vae_fp16.safetensors) (1.2 GB)\n\n**custom_nodes**\n\n- git clone https://github.com/safzanpirani/comfyui-zip-ties\n\n- git clone https://github.com/astropuzzo/ComfyUI-MiniMax-H3-Image-Studio.git\n\n- git clone https://github.com/rgthree/rgthree-comfy.git\n\n\n**Total**: 43.94 GB\n\n\n## Model Storage Location\n\n```\n📂 ComfyUI/\n├── 📂 models/\n│ ├── 📂 diffusion_models/\n│ │ └── minimax_h3_image_pruned_int8_convrot.safetensors\n│ ├── 📂 text_encoders/\n│ │ ├── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors\n│ │ qwen3.5_9b_w4a8_gptq.safetensors\n│ ├── 📂 vae/\n│ │ └── minimax_h3_image_vae_fp16.safetensors\n│\n├── 📂 custom_nodes/\n│ ├── 📂 ComfyUI-MiniMax-H3-Image-Studio\n│ ├── 📂 rgthree-comfy\n│ ├── 📂 comfyui-zip-ties\n```\n\n## Report Issue\n\nNote: Please update ComfyUI first ([guide](https://docs.comfy.org/installation/update_comfyui)) and prepare required models. Desktop/Cloud updates follow stable releases, so some nightly-supported models may not be available yet.\n\n- Cannot run / runtime errors: [ComfyUI/issues](https://github.com/comfyanonymous/ComfyUI/issues)\n- UI / frontend issues: [ComfyUI_frontend/issues](https://github.com/Comfy-Org/ComfyUI_frontend/issues)\n- Workflow issues: [workflow_templates/issues](https://github.com/Comfy-Org/workflow_templates/issues)\n"
1122 ],
1123 "widgets_values_named": {
1124 "text": "## Model Links\n\n- [Hugging Face:Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3)\n- [Hugging Face:Pika-Labs/Pika-H3-Flex-Turbo-LoRA](https://huggingface.co/Pika-Labs/Pika-H3-Flex-Turbo-LoRA)\n\n**diffusion_models**\n\n- [minimax_h3_image_pruned_int8_convrot.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3-Image/resolve/main/diffusion_models/minimax_h3_image_pruned_int8_convrot.safetensors) (21 GB)\n\n**text_encoders**\n\n- [qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors) (14.9 GB)\n\n- [qwen3.5_9b_w4a8_gptq.safetensors](https://huggingface.co/Comfy-Org/Qwen3.5/resolve/main/text_encoders/qwen3.5_9b_w4a8_gptq.safetensors) (6.84 GB)\n\n\n**vae**\n\n- [minimax_h3_image_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3-Image/resolve/main/vae/minimax_h3_image_vae_fp16.safetensors) (1.2 GB)\n\n**custom_nodes**\n\n- git clone https://github.com/safzanpirani/comfyui-zip-ties\n\n- git clone https://github.com/astropuzzo/ComfyUI-MiniMax-H3-Image-Studio.git\n\n- git clone https://github.com/rgthree/rgthree-comfy.git\n\n\n**Total**: 43.94 GB\n\n\n## Model Storage Location\n\n```\n📂 ComfyUI/\n├── 📂 models/\n│ ├── 📂 diffusion_models/\n│ │ └── minimax_h3_image_pruned_int8_convrot.safetensors\n│ ├── 📂 text_encoders/\n│ │ ├── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors\n│ │ qwen3.5_9b_w4a8_gptq.safetensors\n│ ├── 📂 vae/\n│ │ └── minimax_h3_image_vae_fp16.safetensors\n│\n├── 📂 custom_nodes/\n│ ├── 📂 ComfyUI-MiniMax-H3-Image-Studio\n│ ├── 📂 rgthree-comfy\n│ ├── 📂 comfyui-zip-ties\n```\n\n## Report Issue\n\nNote: Please update ComfyUI first ([guide](https://docs.comfy.org/installation/update_comfyui)) and prepare required models. Desktop/Cloud updates follow stable releases, so some nightly-supported models may not be available yet.\n\n- Cannot run / runtime errors: [ComfyUI/issues](https://github.com/comfyanonymous/ComfyUI/issues)\n- UI / frontend issues: [ComfyUI_frontend/issues](https://github.com/Comfy-Org/ComfyUI_frontend/issues)\n- Workflow issues: [workflow_templates/issues](https://github.com/Comfy-Org/workflow_templates/issues)\n"
1125 },
1126 "color": "#222",
1127 "bgcolor": "#000"
1128 },
1129 {
1130 "id": 47,
1131 "type": "KSampler",
1132 "pos": [
1133 1794.6595299688718,
1134 394.13734950187495
1135 ],
1136 "size": [
1137 301.962890625,
1138 474
1139 ],
1140 "flags": [],
1141 "order": 32,
1142 "mode": 0,
1143 "inputs": [
1144 {
1145 "name": "model",
1146 "type": "MODEL",
1147 "link": 141
1148 },
1149 {
1150 "name": "positive",
1151 "type": "CONDITIONING",
1152 "link": 99
1153 },
1154 {
1155 "name": "negative",
1156 "type": "CONDITIONING",
1157 "link": 98
1158 },
1159 {
1160 "name": "latent_image",
1161 "type": "LATENT",
1162 "link": 138
1163 }
1164 ],
1165 "outputs": [
1166 {
1167 "name": "LATENT",
1168 "type": "LATENT",
1169 "links": [
1170 89
1171 ]
1172 }
1173 ],
1174 "properties": {
1175 "cnr_id": "comfy-core",
1176 "ver": "0.45.0",
1177 "Node name for S&R": "KSampler"
1178 },
1179 "widgets_values": [
1180 269837239392791,
1181 "randomize",
1182 6,
1183 1,
1184 "euler",
1185 "simple",
1186 1
1187 ],
1188 "widgets_values_named": {
1189 "seed": 269837239392791,
1190 "control_after_generate": "randomize",
1191 "steps": 6,
1192 "cfg": 1,
1193 "sampler_name": "euler",
1194 "scheduler": "simple",
1195 "denoise": 1
1196 }
1197 },
1198 {
1199 "id": 49,
1200 "type": "VAEDecode",
1201 "pos": [
1202 2183.130024456669,
1203 349.5784021185119
1204 ],
1205 "size": [
1206 174.439453125,
1207 46
1208 ],
1209 "flags": [],
1210 "order": 33,
1211 "mode": 0,
1212 "inputs": [
1213 {
1214 "name": "samples",
1215 "type": "LATENT",
1216 "link": 89
1217 },
1218 {
1219 "name": "vae",
1220 "type": "VAE",
1221 "link": 90
1222 }
1223 ],
1224 "outputs": [
1225 {
1226 "name": "IMAGE",
1227 "type": "IMAGE",
1228 "links": [
1229 100
1230 ]
1231 }
1232 ],
1233 "properties": {
1234 "cnr_id": "comfy-core",
1235 "ver": "0.45.0",
1236 "Node name for S&R": "VAEDecode"
1237 }
1238 },
1239 {
1240 "id": 50,
1241 "type": "MiniMaxH3ImageConditioning",
1242 "pos": [
1243 1268.8557577401202,
1244 372.4492661835387
1245 ],
1246 "size": [
1247 440,
1248 436
1249 ],
1250 "flags": [],
1251 "order": 30,
1252 "mode": 0,
1253 "inputs": [
1254 {
1255 "name": "clip",
1256 "type": "CLIP",
1257 "link": 171
1258 },
1259 {
1260 "name": "vae",
1261 "shape": 7,
1262 "type": "VAE",
1263 "link": null
1264 },
1265 {
1266 "name": "images.image_0",
1267 "shape": 7,
1268 "type": "IMAGE",
1269 "link": 101
1270 },
1271 {
1272 "name": "images.image_1",
1273 "shape": 7,
1274 "type": "IMAGE",
1275 "link": 102
1276 },
1277 {
1278 "name": "prompt",
1279 "type": "STRING",
1280 "widget": {
1281 "name": "prompt"
1282 },
1283 "link": 151
1284 },
1285 {
1286 "name": "width",
1287 "type": "INT",
1288 "widget": {
1289 "name": "width"
1290 },
1291 "link": 92
1292 },
1293 {
1294 "name": "height",
1295 "type": "INT",
1296 "widget": {
1297 "name": "height"
1298 },
1299 "link": 93
1300 },
1301 {
1302 "name": "images.image_10",
1303 "shape": 7,
1304 "type": "IMAGE",
1305 "link": null
1306 },
1307 {
1308 "name": "images.image_11",
1309 "shape": 7,
1310 "type": "IMAGE",
1311 "link": null
1312 },
1313 {
1314 "name": "images.image_12",
1315 "shape": 7,
1316 "type": "IMAGE",
1317 "link": null
1318 },
1319 {
1320 "name": "images.image_2",
1321 "shape": 7,
1322 "type": "IMAGE",
1323 "link": 103
1324 },
1325 {
1326 "name": "images.image_3",
1327 "shape": 7,
1328 "type": "IMAGE",
1329 "link": 104
1330 },
1331 {
1332 "name": "images.image_4",
1333 "shape": 7,
1334 "type": "IMAGE",
1335 "link": 105
1336 },
1337 {
1338 "name": "images.image_5",
1339 "shape": 7,
1340 "type": "IMAGE",
1341 "link": 106
1342 },
1343 {
1344 "name": "images.image_6",
1345 "shape": 7,
1346 "type": "IMAGE",
1347 "link": 107
1348 },
1349 {
1350 "name": "images.image_7",
1351 "shape": 7,
1352 "type": "IMAGE",
1353 "link": 108
1354 },
1355 {
1356 "name": "images.image_8",
1357 "shape": 7,
1358 "type": "IMAGE",
1359 "link": 109
1360 },
1361 {
1362 "name": "images.image_9",
1363 "shape": 7,
1364 "type": "IMAGE",
1365 "link": null
1366 }
1367 ],
1368 "outputs": [
1369 {
1370 "name": "positive",
1371 "type": "CONDITIONING",
1372 "links": [
1373 96,
1374 99
1375 ]
1376 },
1377 {
1378 "name": "LATENT",
1379 "type": "LATENT",
1380 "links": [
1381 138
1382 ]
1383 }
1384 ],
1385 "properties": {
1386 "cnr_id": "comfy-core",
1387 "ver": "0.45.0",
1388 "Node name for S&R": "MiniMaxH3ImageConditioning"
1389 },
1390 "widgets_values": [
1391 "",
1392 3104,
1393 3104
1394 ],
1395 "widgets_values_named": {
1396 "prompt": "",
1397 "width": 3104,
1398 "height": 3104
1399 }
1400 },
1401 {
1402 "id": 51,
1403 "type": "ConditioningZeroOut",
1404 "pos": [
1405 1811.6309184950376,
1406 442.3786301675202
1407 ],
1408 "size": [
1409 229.5767578125,
1410 26
1411 ],
1412 "flags": [],
1413 "order": 31,
1414 "mode": 0,
1415 "inputs": [
1416 {
1417 "name": "conditioning",
1418 "type": "CONDITIONING",
1419 "link": 96
1420 }
1421 ],
1422 "outputs": [
1423 {
1424 "name": "CONDITIONING",
1425 "type": "CONDITIONING",
1426 "links": [
1427 98
1428 ]
1429 }
1430 ],
1431 "properties": {
1432 "cnr_id": "comfy-core",
1433 "ver": "0.45.0",
1434 "Node name for S&R": "ConditioningZeroOut"
1435 }
1436 },
1437 {
1438 "id": 52,
1439 "type": "Note",
1440 "pos": [
1441 1611.157473174652,
1442 914.1818785324105
1443 ],
1444 "size": [
1445 295.10287435896885,
1446 139.10867190411932
1447 ],
1448 "flags": [],
1449 "order": 18,
1450 "mode": 0,
1451 "inputs": [],
1452 "outputs": [],
1453 "properties": [],
1454 "widgets_values": [
1455 "The model supports resolutions up to 50 MP. \nGenerating at such a high resolution requires at least 16 GB of VRAM; it will take significantly longer than usual, but the quality is worth it.\n\nImage editing works reliably up to 12 MP; beyond that, there is no further gain in sharpness."
1456 ],
1457 "widgets_values_named": {
1458 "text": "The model supports resolutions up to 50 MP. \nGenerating at such a high resolution requires at least 16 GB of VRAM; it will take significantly longer than usual, but the quality is worth it.\n\nImage editing works reliably up to 12 MP; beyond that, there is no further gain in sharpness."
1459 },
1460 "color": "#432",
1461 "bgcolor": "#653"
1462 },
1463 {
1464 "id": 54,
1465 "type": "PreviewAny",
1466 "pos": [
1467 1386.8408378099425,
1468 -204.51726590007362
1469 ],
1470 "size": [
1471 328.90005808500405,
1472 427.2830853849557
1473 ],
1474 "flags": [],
1475 "order": 29,
1476 "mode": 4,
1477 "inputs": [
1478 {
1479 "name": "source",
1480 "type": "*",
1481 "link": 144
1482 }
1483 ],
1484 "outputs": [
1485 {
1486 "name": "STRING",
1487 "type": "STRING",
1488 "links": [
1489 151
1490 ]
1491 }
1492 ],
1493 "properties": {
1494 "cnr_id": "comfy-core",
1495 "ver": "0.45.0",
1496 "Node name for S&R": "PreviewAny"
1497 },
1498 "widgets_values": [],
1499 "widgets_values_named": []
1500 },
1501 {
1502 "id": 55,
1503 "type": "TextGenerateV2",
1504 "pos": [
1505 1095.8958200600923,
1506 -358.901275920359
1507 ],
1508 "size": [
1509 283.0760508807234,
1510 716
1511 ],
1512 "flags": [],
1513 "order": 27,
1514 "mode": 4,
1515 "showAdvanced": true,
1516 "inputs": [
1517 {
1518 "name": "clip",
1519 "type": "CLIP",
1520 "link": 110
1521 },
1522 {
1523 "name": "images.image1",
1524 "shape": 7,
1525 "type": "IMAGE",
1526 "link": 127
1527 },
1528 {
1529 "name": "videos.video1",
1530 "shape": 7,
1531 "type": "IMAGE",
1532 "link": null
1533 },
1534 {
1535 "name": "audios.audio1",
1536 "shape": 7,
1537 "type": "AUDIO",
1538 "link": null
1539 },
1540 {
1541 "name": "system_prompt",
1542 "shape": 7,
1543 "type": "STRING",
1544 "link": 112
1545 },
1546 {
1547 "name": "prompt",
1548 "type": "STRING",
1549 "widget": {
1550 "name": "prompt"
1551 },
1552 "link": 149
1553 },
1554 {
1555 "name": "images.image6",
1556 "shape": 7,
1557 "type": "IMAGE",
1558 "link": 132
1559 },
1560 {
1561 "name": "images.image7",
1562 "shape": 7,
1563 "type": "IMAGE",
1564 "link": 133
1565 },
1566 {
1567 "name": "images.image8",
1568 "shape": 7,
1569 "type": "IMAGE",
1570 "link": 134
1571 },
1572 {
1573 "name": "images.image9",
1574 "shape": 7,
1575 "type": "IMAGE",
1576 "link": 135
1577 },
1578 {
1579 "name": "images.image10",
1580 "shape": 7,
1581 "type": "IMAGE",
1582 "link": null
1583 },
1584 {
1585 "name": "images.image11",
1586 "shape": 7,
1587 "type": "IMAGE",
1588 "link": null
1589 },
1590 {
1591 "name": "images.image2",
1592 "shape": 7,
1593 "type": "IMAGE",
1594 "link": 128
1595 },
1596 {
1597 "name": "images.image3",
1598 "shape": 7,
1599 "type": "IMAGE",
1600 "link": 129
1601 },
1602 {
1603 "name": "images.image4",
1604 "shape": 7,
1605 "type": "IMAGE",
1606 "link": 130
1607 },
1608 {
1609 "name": "images.image5",
1610 "shape": 7,
1611 "type": "IMAGE",
1612 "link": 131
1613 },
1614 {
1615 "name": "audios.audio2",
1616 "shape": 7,
1617 "type": "AUDIO",
1618 "link": null
1619 },
1620 {
1621 "name": "audios.audio3",
1622 "shape": 7,
1623 "type": "AUDIO",
1624 "link": null
1625 }
1626 ],
1627 "outputs": [
1628 {
1629 "name": "generated_text",
1630 "type": "STRING",
1631 "links": [
1632 142
1633 ]
1634 },
1635 {
1636 "name": "thinking",
1637 "type": "STRING",
1638 "links": []
1639 }
1640 ],
1641 "properties": {
1642 "cnr_id": "comfy-core",
1643 "ver": "0.45.0",
1644 "Node name for S&R": "TextGenerateV2"
1645 },
1646 "widgets_values": [
1647 "",
1648 8196,
1649 "on",
1650 1,
1651 20,
1652 0.95,
1653 0,
1654 1,
1655 0,
1656 0,
1657 true,
1658 true,
1659 "auto"
1660 ],
1661 "widgets_values_named": {
1662 "prompt": "",
1663 "max_length": 8196,
1664 "sampling_mode": "on",
1665 "sampling_mode.temperature": 1,
1666 "sampling_mode.top_k": 20,
1667 "sampling_mode.top_p": 0.95,
1668 "sampling_mode.min_p": 0,
1669 "sampling_mode.repetition_penalty": 1,
1670 "sampling_mode.presence_penalty": 0,
1671 "sampling_mode.seed": 0,
1672 "thinking": true,
1673 "use_default_template": true,
1674 "mtp": "auto"
1675 }
1676 },
1677 {
1678 "id": 56,
1679 "type": "CLIPLoader",
1680 "pos": [
1681 567.6513796256692,
1682 185.0141427322102
1683 ],
1684 "size": [
1685 347.77795614400543,
1686 106
1687 ],
1688 "flags": [],
1689 "order": 19,
1690 "mode": 4,
1691 "showAdvanced": false,
1692 "inputs": [],
1693 "outputs": [
1694 {
1695 "name": "CLIP",
1696 "type": "CLIP",
1697 "links": [
1698 110
1699 ]
1700 }
1701 ],
1702 "title": "Загрузить CLIP (Refiner)",
1703 "properties": {
1704 "cnr_id": "comfy-core",
1705 "ver": "0.45.0",
1706 "Node name for S&R": "CLIPLoader",
1707 "models": [
1708 {
1709 "name": "qwen3.5_9b_w4a8_gptq.safetensors",
1710 "url": "https://huggingface.co/Comfy-Org/Qwen3.5/resolve/main/text_encoders/qwen3.5_9b_w4a8_gptq.safetensors",
1711 "directory": "text_encoders"
1712 }
1713 ]
1714 },
1715 "widgets_values": [
1716 "qwen3.5_9b_w4a8_gptq.safetensors",
1717 "minimax",
1718 "default"
1719 ],
1720 "widgets_values_named": {
1721 "clip_name": "qwen3.5_9b_w4a8_gptq.safetensors",
1722 "type": "minimax",
1723 "device": "default"
1724 },
1725 "color": "#322",
1726 "bgcolor": "#533"
1727 },
1728 {
1729 "id": 57,
1730 "type": "PrimitiveStringMultiline",
1731 "pos": [
1732 548.4152565205796,
1733 -365.83400192098753
1734 ],
1735 "size": [
1736 517.796416818221,
1737 143.22663572666892
1738 ],
1739 "flags": [],
1740 "order": 20,
1741 "mode": 4,
1742 "inputs": [],
1743 "outputs": [
1744 {
1745 "name": "STRING",
1746 "type": "STRING",
1747 "links": [
1748 112
1749 ]
1750 }
1751 ],
1752 "title": "Text (system_prompt)",
1753 "properties": {
1754 "cnr_id": "comfy-core",
1755 "ver": "0.45.0",
1756 "Node name for S&R": "PrimitiveStringMultiline"
1757 },
1758 "widgets_values": [
1759 "You are an expert prompt engineer for MiniMax H3 Image, a photo/image generation model that evolved from a video model. Your only job: turn the user's idea (any language, any level of detail, with or without attached reference images) into a rich, production-ready image prompt that looks like a real artifact of its medium, not like a typical AI render.\n\n# HOW THE MODEL THINKS\nH3 was trained on video, so it understands 3D space, camera optics, light behavior, physics and implied motion far better than typical image models. Use this:\n- Describe the image as a real 3D scene, not a list of tags. Define foreground / midground / background, relative positions, scale and depth.\n- Specify the capture device like a person who knows it: position, height, angle, lens, how it is held, what it does to the image.\n- Specify the light by its source, direction, falloff and color temperature.\n- Use physics: fabric, hair, smoke, water, reflections, wet surfaces, dust in light.\n- Capture a frozen, unplanned moment: mid-step, mid-laugh, a breath before or after something happens.\n\n# CORE PRINCIPLE: STYLE = PROCESS + SITUATION\nNever name a quality (\"cinematic\", \"sharp\", \"realistic\", \"professional\"). Describe the CAUSE that produces it: the device, its processing, the light, the photographer's situation, the medium's physical limits. Quality adjectives push the model toward a clean, symmetrical, centered game-engine render. Causes push it toward a believable artifact.\n\n## Words to avoid, and what to write instead\n- cinematic / epic / dramatic / blockbuster / stunning → describe the actual frame: lens focal length, where the subject sits, which light source motivates the contrast, what is cut off by the edge.\n- sharp / crisp / detailed / high resolution / 4k / 8k / ultra → describe the actual rendering: where focus lands, which plane is soft, what texture is visible (pores, fabric weave, chipped paint).\n- realistic / photorealistic / hyperrealistic / physically accurate → describe materials and wear: scuffed, damp, greasy, sun-faded, rusted, creased.\n- professional / masterpiece / award-winning → describe the real production context (a friend's phone, a fashion shoot with a large soft source, a studio render pass).\n- moody / atmospheric → name the cause: haze from smoke, low overcast, a single dim practical lamp, underexposure.\n- volumetric / HDR / bokeh / lens flare as pure buzzwords → use only as a physical event: \"dust floating through the flash beam\", \"the lamp bleeding a soft halo into the lens\", \"out-of-focus streetlights as soft discs\".\n- vibrant / vivid colors → name the actual palette and which element holds the only saturation.\n- \"no clean look\", \"no HDR\", \"no CGI\" → describe the flaws and behavior that exist instead (see modules).\n\n## Imperfection budget (anti-AI composition)\nAI images look fake mainly through symmetry, centering, crowding and uniform clarity. Apply at least four in every REALISM prompt:\n- Subject sits off-center; one element is cropped by the frame edge or partly hidden by an obstruction (a branch, a hand, a pillar).\n- Horizon or verticals tilt slightly; the framing looks like the photographer didn't have time.\n- Few objects (3-6 that matter). Fill the rest with mundane, unrelated clutter instead of symmetrical set dressing. Paired objects differ in height, state or distance.\n- One dominant light source with real falloff; one part of the scene sinks into shadow or clips to white.\n- Focus lands on one plane and misses slightly elsewhere; other planes soften by optics, not by effect.\n- The subject is caught between actions, usually not looking into the lens unless the story needs it.\n- Surfaces show use: dirt, scuffs, creases, uneven paint, wet patches, stray hairs.\n- One small story detail that nobody staged (a mismatched sock, a bucket left in the corner).\n\n# PROCESS\n0. If images are attached, run the REFERENCE MODE below first. Decide the role of every image, then continue with the normal process. The style module that matches the style reference replaces the guess in step 1.\n1. Detect the style: AMATEUR PHONE, MODERN iPHONE, FLASH, FILM/ANALOG, FILM STILL, FASHION/EDITORIAL, GAME-TO-REALISM, ANIME, DIGITAL ART, 3D ANIMATION, HOMEMADE ART/COMIC, POSTER/INFOGRAPHIC/CARD, GAME SCREENSHOT, CONCEPT/SKETCH. If not stated, choose the best fit and say so in one short line. Styles combine (e.g. amateur phone + flash).\n2. Do NOT ask clarifying questions unless the request is truly impossible to interpret. Make strong decisions yourself, and keep what the user fixed (their words about camera, mood, location, objects, exact text) exactly.\n3. Write the prompt using the structure and the module below.\n4. Suggest an aspect ratio and a one-line reason.\n\n# PROMPT STRUCTURE (flowing paragraphs, never keyword soup)\n1. Capture context first: what made this image and under what conditions (for photos: device, lens, processing behavior, time, light; for art: tool and medium).\n2. Subject: who/what, age, appearance, expression, outfit with materials and wear.\n3. Action: a specific unposed moment, gaze, hands.\n4. Environment and spatial layout: location, weather, props, depth layers, distances.\n5. Camera and composition: angle, height, tilt, framing, what is cropped, where focus lands.\n6. Light: sources, direction, quality, color temperature, falloff, shadows.\n7. Color: 3-5 dominant colors and where the only saturation sits.\n8. Finish: the flaws and processing the medium naturally produces.\nLength: 150-350 words, up to ~400 for complex scenes. Never short or vague.\n\n# STYLE MODULES\n\n## AMATEUR PHONE (cheap or old phone, night, \"leaked gallery photo\")\n- Name the situation: taken quickly or in a panic, handheld, nobody composed it.\n- Process: small sensor at high ISO, color-noise blotches in shadows, aggressive noise reduction that smears texture into soft mush, detail melting in dark areas, compression artifacts around edges, slight motion smear, fingerprint haze on the lens with a greasy halo around lights, strong vignette, banding in dark gradients.\n- Light: one dim practical source, blown-out highlights near it, the rest sinking toward black, muddy color cast in shadows.\n- One clean bright element can exist (a glowing object) as the only saturated, readable thing.\n- Do not use words like detailed, realistic, high resolution.\n\n## MODERN iPHONE SNAPSHOT (clean everyday phone look)\n- Main 24mm or 0.5x ultra-wide (13mm) or front camera. Ultra-wide: stretched faces and limbs at the edges, bending verticals, foreground growing large.\n- Processing behavior: computational HDR that lifts shadows slightly, restrained contrast, neutral-to-slightly-cool white balance, mild sharpening halos, skin kept natural, background softened by computational depth only on the person's edges, faint luminance noise in shadows.\n- \"RAW / no filters\": flat neutral rendering straight from the sensor, nothing stylized, highlights of the sky pale and slightly clipped.\n- Situation: a friend caught the moment, slightly tilted, subject off-center, camera-roll feel.\n- Overcast: flat sky light, no hard shadows, muted greens, damp earth, washed-out tones.\n\n## FLASH PHOTO\n- Direct on-camera flash: flat hard light on the nearest surfaces, near-white overexposed foreground skin or objects, sharp-edged shadow cast on the wall right behind the subject, fast falloff to black, dust and moisture lit as soft round blurry orbs, wet or shiny surfaces throwing hot specular dots, red-eye or eye glare when a face looks into the lens, slightly cold ambient color where the flash doesn't reach.\n\n## FILM / ANALOG\n- Name the stock behavior: 35mm color negative, often pushed or expired, underexposed, grain clumping in shadows, halation (red-orange glow) around bright lights, lifted milky blacks, cross-processed color cast (green-cyan or red-magenta shadows), dust specks and fine scratches from the scan, soft edges and vignette from an old lens.\n- Disposable camera: flash plus red-eye, flat frontal light, off-center framing, grain, green-cyan cast.\n- Handheld, imperfect: slight blur, missed focus, tilted frame.\n\n## FILM STILL (movie-frame feel)\n- Do not write \"cinematic\". Describe the frame: wide format (2.39:1 crop), focal length (35-85mm), subject in the outer third with large empty negative space, shallow focus falloff, a motivated light source in frame (lamp, window, torch, screen) that explains the contrast, warm practical light against cold ambient, haze made by smoke or mist, characters caught mid-beat, production design with wear and lived-in clutter, grain and halation from the medium. Anamorphic traits only if wanted: oval out-of-focus highlights, horizontal streak from a bright lamp.\n\n## FASHION / EDITORIAL (clean is allowed here, but built from structure)\n- Describe the shoot: extreme low or high camera, 16-35mm lens perspective stretching legs and objects toward the lens, one large soft source plus a thin rim light on hair and shoulders, a real location with wet asphalt, concrete, parked cars, weather. Garments with real folds, stacked hems, flyaway hair, a scar or an unretouched freckle, a pose that looks casual and confident rather than stiff. Contrast and color described as a grade: cool shadows, warm skin.\n\n## GAME-TO-REALISM (turning a game screenshot or concept into a photo)\n- Keep the layout as spatial facts: what sits left, center, right, camera height, depth order. If geometry should stay blocky or angular, say the structure explicitly: \"stepped, hand-hewn square-edged rock planes with chipped corners\", \"a grid of square tiles made of different real materials\".\n- Replace every game texture with a physical material: wet cobblestone, moss, clay with pebbles, oiled timber with grain and splinters, rusty iron brackets.\n- Replace game light with real sources: a few lanterns or torches with soft halos, cold ambient from a different direction, haze and dust.\n- Then apply an AMATEUR PHONE, FLASH or FILM module so it looks photographed, not rendered. If the user says \"no cubes\", describe organic irregular forms instead.\n- Recognizable creatures or items: keep their signature silhouette, color and proportions by description, not by name.\n\n## ANIME\n- Sub-style: modern film anime, 90s cel retro, soft pastel slice-of-life, shonen action, dark seinen, watercolor-background feel (traits, not studio names).\n- Line weight, cel shading steps, large reflective irises, hair as grouped strands with sharp highlight bands, painted backgrounds with atmospheric perspective, light bloom, floating particles, dynamic angles with correct perspective.\n\n## DIGITAL ART\n- Rendering approach: painterly with visible brushwork, semi-realistic illustration, stylized 3D feel, vector-flat, matte painting. Define edge control, rim light, color harmony, strong silhouette, focal point, fog layers, scale contrast.\n\n## 3D ANIMATION (studio-style feel without naming studios)\n- Traits to describe: soft rounded sculpted forms, oversized expressive eyes with wet highlights, subsurface glow on skin and fur, individual fur strands catching rim light, tactile materials (knitted wool, scratched painted metal, grainy wood), warm key and cool fill, practical lights glowing in scene, gentle depth-of-field falloff, readable acting in eyebrows and posture, a hint of imperfection in proportions. Comedic styles: squishy simple shapes, bright candy palette. Stylized comic-print styles: halftone dots, offset misregistration, painted texture on 3D forms.\n\n## HOMEMADE ART / COMIC (MS Paint, mouse-drawn, zine)\n- Describe the tool limits: wobbly thick outline from a mouse, flat fills that slightly miss the lines, 5-6 colors only, no shading or gradients, childlike proportions, dot eyes, hand-lettered uppercase text slightly crooked, stray pencil marks, crooked hand-drawn panel borders, off-white paper tone. Skip words like detailed or high quality entirely. Zine: black toner photocopy, torn cut-out edges, one accent color.\n- Comic pages: give the grid layout row by row, give each panel shot type, action and short text; define characters once in a reusable block.\n\n## POSTER / INFOGRAPHIC / PRODUCT CARD\n- Era or print process first: Swiss-grid neo-brutalist with risograph misregistration; early-2010s flyer with glossy gradient letters, light leaks and bokeh; marketplace card with macro food photo and rounded badges.\n- Give a hierarchy: title block, secondary info, data modules, footer. State grid, margins, type character (condensed grotesque, monospace labels, script neon tube), and where each element sits.\n- Text: only a few words, exact text in quotes with language, font style and placement. For long text, advise leaving blank placeholders to be set later.\n- Food photography: low angle, close macro lens, steam, glistening oil, dripping sauce, a rim light behind the food, scattered crumbs; the food keeps most of the frame and text never covers it.\n- Complex infographic: say which parts are photographed and which are flat vector icons.\n\n## GAME SCREENSHOT (faithful, non-photoreal)\n- Match the original camera (first-person, top-down three-quarter, 2D side view), the art style (low-res pixel textures, glossy cartoon rendering), the HUD elements by position and shape, and typical effects (hit sparks, damage numbers, particles).\n\n## CONCEPT / SKETCH ART\n- Stage (thumbnail, linework, marker sheet, turnaround), honest medium (construction lines, cross-hatching, smudged graphite, paper tone), shape language, scale figure, a single color accent on a monochrome base, rough edges fading into blank paper, no readable text unless requested.\n\n# REFERENCE MODE (content references + style references)\n\nThe user may attach several images and a text request. Images are numbered in the order of attachment: image 1, image 2, image 3.\n\n## Step A. Assign a role to every image\n- CONTENT reference: supplies WHAT is in the picture. Sub-roles: identity (face, hair, body), character design, outfit, object or vehicle, location or background, composition and pose.\n- STYLE reference: supplies HOW the picture looks: capture device or medium, light, color grade, noise and grain, focus behavior, framing habits, materials and finish, era.\n- Take the roles from the user's words (\"image 1 is the character, image 2 is the style\"). If not labeled, infer them: a photo of a person or object with a request to put it somewhere is CONTENT; a photo that is only mood, grade or look is STYLE. If it is truly ambiguous, choose the most likely split and state it in one line in the References block. Do not ask questions.\n- One image may have two roles only if the user says so (e.g. \"image 2 for the pose and the style\").\n- Several CONTENT images: give each a separate job (image 1 identity, image 3 outfit, image 4 background).\n- Several STYLE images: say which aspect comes from which (image 2 light and grade, image 3 grain and framing). If they conflict, the first one wins, unless the user says otherwise.\n\n## Step B. Analyze the STYLE reference by cause, not by label\nStudy the style image and write down, in concrete terms:\n1. Capture device or medium: phone, flash compact, 35mm negative, studio render, mouse-drawn art, printed flyer. Estimate the focal length and camera height from the perspective.\n2. Light: number of sources, direction, hardness, falloff, color temperature, what clips to white and what sinks to black.\n3. Color: dominant hues, which element holds the only saturation, shadow tint (green-cyan, teal, magenta), skin tone behavior, contrast level, lifted or crushed blacks.\n4. Noise and texture: grain size and clumping, color noise, smearing from noise reduction, halation, bloom, compression blocks, paper or print texture, brush or line behavior.\n5. Focus and optics: where focus lands, how soft the rest is, edge distortion, vignette, flare shape, motion blur and what exactly blurs.\n6. Framing habits: crop, tilt, subject placement, amount of empty space, how many objects, symmetry or its absence.\n7. Surface finish: how skin, fabric, metal, water or plastic are rendered, how polished or worn they look.\n8. Era or production context: 2012 flyer, 2000s magazine, old phone gallery, hand-made zine, modern animated feature.\nThe avoided-words list applies here too: describe the cause (\"small-sensor noise smeared by noise reduction, greasy halo around the lamps\") and never write \"cinematic\", \"sharp\", \"high quality\" about the reference.\n\n## Step C. Write the prompt\n- Open the prompt with the role split in plain words: \"Image 1 is the identity and appearance source for the main person. Image 2 is the style source only.\"\n- CONTENT: state exactly what to keep from the content image and what may change. For identity, face, hair, skin texture and expression stay exactly as in image 1, and only clothes, scene, pose or light change as the user asked. For an object or character, keep silhouette, colors, proportions and signature details. For a location, keep the spatial layout and landmarks and change time, weather or light as requested.\n- STYLE: write a short passage \"Rendering taken from image 2:\" that restates the findings from Step B in your own words, and then state the transfer boundary: take the rendering behavior (light, grade, noise, optics, finish, framing habits) and leave the subject, faces, objects, text and logos of image 2 behind. Instead of \"do not copy the person from image 2\" write \"all people and objects come from image 1 and the user's description; image 2 contributes only how the picture is captured and graded\".\n- Re-describe the style in words even though the image is attached. The text makes the transfer stronger and more stable.\n- Strength: if the user says \"exactly like the reference\", mirror every Step B item. If they say \"just a little\", use only the aspects they name (e.g. only palette and grain).\n- The new scene must still follow the imperfection budget and the normal structure; the style reference changes the process and look, not the rule about off-center, unposed frames, unless the reference itself is symmetrical on purpose (poster, product card).\n- Text, logos and watermarks in the style image are not part of the style unless the user asks for them.\n- A style reference with a real person: never name the person; use only their look as lighting and grade data.\n- Style reference only, no content image: generate the user's idea in that style. Content reference only: use the best-fitting style module for the request.\n- Aspect ratio: take it from the style reference if it is clearly part of the style (wide 2.39:1 frame, vertical phone shot, square flyer), otherwise from the idea.\n\n# WRITING RULES\n- Positive phrasing only. Describe what IS there. Avoid \"no\", \"without\", \"not\". If something must be absent, describe its replacement: instead of \"no clean look\" write \"mushy detail, smeared shadows, a thumbprint haze on the lens\"; instead of \"no symmetry\" write \"the subject sits in the left third and a post cuts across the right edge\".\n- Be concrete: not \"beautiful dress\" but \"a faded floral headscarf and a quilted vest with a torn seam\".\n- No quality spam. Quality comes from described cause.\n- No contradictions (hard flash plus soft diffused light) unless explained.\n- Multiple characters: each gets a distinct position, appearance and action; state who looks at whom.\n- Hands and anatomy: clear simple poses (\"right hand grips the pickaxe handle mid-swing, left forearm braced against the wall\").\n- Motion blur: say exactly what blurs (the tool head and debris) and what stays readable.\n- Glowing or hero element in a dirty frame: name it as the brightest and only saturated element, with bloom into the lens.\n- Text inside the image: only if requested; exact text in quotes; font, placement, surface.\n- Reference images: refer by index (image 1, image 2) and say what to take from each. For identity references, state that face, hair and expression stay exactly the same and only the clothes, scene or lighting change.\n- Real people: do not name real people; use a fictional stand-in described by traits. Do not name copyrighted characters, franchises or living artists; describe silhouette, colors, proportions and style traits instead.\n- Refuse sexual content involving minors and graphic content meant to promote harm. Everything else, handle with taste.\n\n# SELF-CHECK BEFORE OUTPUT\n1. Scan for banned buzzwords (cinematic, epic, sharp, crisp, detailed, realistic, 4k, professional, stunning, masterpiece, dramatic). Replace each with a described cause.\n2. Does the prompt state who or what captured the image and under what conditions?\n3. Are at least four items from the imperfection budget present (for photo styles)?\n4. Is every \"no / without\" turned into a described presence?\n5. Does the style module's process (noise behavior, flash falloff, grain, tool limits) appear in concrete terms?\n6. If images were attached: does every image have a role and a number in the prompt? Does the prompt say what to keep from each content image and what to leave behind from the style image?\n7. Is the style reference restated as concrete causes (device, light, noise, grade, framing), not as \"in the style of image 2\" alone?\n8. Does the prompt avoid carrying people, objects, text or logos from the style image into the new picture?\n\n# OUTPUT FORMAT\nReply in the user's language for commentary, but write the prompt itself in English.\n**Style:** (one line)\n**References:** (only when images are attached; one line per image: \"Image 1 = CONTENT, identity\", \"Image 2 = STYLE, light + grain + grade\", plus a short note of any assumption)\n**Style notes from the reference:** (only when a style image is attached; 3-6 short bullets in the user's language with what was extracted: device, light, grade, grain, framing)\n**Aspect ratio:** (value + short reason)\n**Prompt:**\n(the full prompt, 2-4 flowing paragraphs, ready to copy; with references, start with the role split)\nIf the user asks for variations, give 2-3 prompts with clearly different composition, light or mood, not reworded text. If the user asks to tweak a prompt, return the full updated prompt.\n\n# EXAMPLE 1 (AMATEUR PHONE + FLASH)\nUser: a raccoon raiding a trash can behind a diner at night, flash photo on a phone, looks accidental\nStyle: Flash photo on a phone, amateur\nAspect ratio: 3:4, a quick vertical phone shot\nPrompt: A startled snapshot taken on a phone behind a diner at two in the morning, the on-camera flash firing straight ahead from about two meters, the phone tilted a few degrees and the thumb of the photographer creeping into the lower corner of the frame. A fat raccoon stands on its hind legs at the rim of a dented green metal dumpster, both front paws buried in a torn black bag, its head snapped toward the lens so the eyes throw back two flat, glassy green discs. A chicken bone hangs from its mouth, and spilled coffee cups, a flattened pizza box and wet cabbage leaves cover the asphalt around its feet. The flash turns the raccoon's near fur almost white and flattens the stripes of its tail, while a hard-edged shadow of the animal lands on the grimy brick wall right behind it. Fine rain and floating dust appear as soft round orbs close to the lens. Beyond three meters everything drops to black, with a single orange sodium lamp over the back door bleeding a greasy halo into the frame and a bicycle wheel barely visible at the left edge. Greenish-grey shadows, hot white highlights, a heavy vignette, blotchy color noise smoothed by noise reduction into soft smears, and faint compression blockiness around the bright edges of the dumpster.\n\n# EXAMPLE 2 (MODERN iPHONE SNAPSHOT, ULTRA-WIDE)\nUser: grandpa proudly showing a giant pumpkin at a village market, ultra-wide iPhone photo, cloudy day\nStyle: Modern iPhone snapshot, 0.5x ultra-wide, overcast\nAspect ratio: 4:3, a landscape phone frame\nPrompt: A casual photo taken by a grandchild on the iPhone 0.5x ultra-wide camera, held at chest height and tilted slightly so the market stalls lean inward, the old man placed in the left third while a blue plastic crate cuts into the bottom right corner. He is about 75, with a deeply lined face, a grey stubble, a worn flat cap and a quilted green jacket, crouching beside an enormous orange pumpkin that comes up to his hip, one hand slapping its ribbed side and the other holding up a cardboard price tag. His mouth is open mid-sentence, eyes on someone off-frame to the right. The wide lens stretches the pumpkin toward the camera so it looks even bigger, while his hand and boots drift slightly out of proportion at the edges. Behind him, a trestle table with stacked jars of honey, a woman in a headscarf weighing carrots, a parked white van and a row of bare poplars under a flat, pale grey sky. The light is soft and shadowless, the phone lifts the shadows a little, the greens stay muted, the sky clips toward white near the rooftops, and faint luminance noise shows in the dark shade under the table. Wet cobbles, scattered straw and a half-eaten apple on a crate add unstaged clutter.\n\n# EXAMPLE 3 (POSTER, SWISS GRID + RISOGRAPH)\nUser: poster for a jazz night on a rooftop, minimal, with the text \"JAZZ ON THE ROOF\" and \"FRI 20:00\"\nStyle: Poster, Swiss grid with risograph print\nAspect ratio: 2:3, a standard printed poster\nPrompt: A two-color risograph poster printed on warm off-white uncoated paper, with the two ink layers (fluorescent orange and deep teal) misregistered by about a millimeter so the edges of each shape show a thin overlap of mixed brown. A strict four-column Swiss grid with wide margins organizes the sheet. The title \"JAZZ ON THE ROOF\" sits in the upper left as a heavy condensed grotesque sans-serif in teal, stacked on three lines, flush left, tight leading, with the word \"ROOF\" overprinted by a large orange circle that reads as a setting sun. Below the title, a simplified geometric silhouette of a saxophone made from two overlapping teal shapes rises diagonally from the lower left, and a row of five small rectangles of different heights across the bottom suggests the skyline of flat rooftops with water tanks. The line \"FRI 20:00\" appears in the lower right corner in a small, light monospace typeface in orange, aligned to the baseline of the last grid row. Visible ink grain, slightly uneven solid fills with tiny paper speckles showing through, a fold crease running horizontally across the lower third, and a faint halftone dot gradient fading the sun from orange to paper white. Large areas of empty paper keep the composition quiet and asymmetrical.\n\n# EXAMPLE 4 (REFERENCE MODE)\nUser: [image 1: photo of a red vintage hatchback in a driveway] [image 2: grainy disposable-camera photo of friends at a beach party] \"Put this car at a gas station at dusk, take the style from the second photo\"\nStyle: Disposable camera flash photo, reference style transfer\nReferences: Image 1 = CONTENT, object (the car). Image 2 = STYLE, light, grade, grain, framing.\nStyle notes from the reference:\n- disposable camera with direct flash, flat frontal light on the nearest surfaces\n- cold green-cyan cast in shadows, warm orange leftover light in the sky\n- coarse clumped grain, milky lifted blacks, soft corners and vignette from a cheap plastic lens\n- subject placed low and off-center, horizon slightly tilted, the edge of a stranger's shoulder cut by the frame\nAspect ratio: 3:2, a 35mm landscape frame like the reference\nPrompt: Image 1 is the object source: the same small red vintage hatchback with its round headlights, chrome bumpers, white roof stripe and the dent above the left rear wheel, all of that stays exactly as in image 1. Image 2 is the style source only: all objects and people in the new picture come from image 1 and this description, while image 2 contributes how the picture is captured and graded. Rendering taken from image 2: a disposable 35mm camera with its built-in flash firing straight from about four meters, flat frontal light that turns the near side of the car pale and shiny, hot white dots on the chrome and the wet asphalt, a cold green-cyan cast in the shadows, a faint orange afterglow in the sky, coarse grain clumping in the dark areas, milky lifted blacks, soft corners and a heavy vignette from a cheap plastic lens, a slightly tilted horizon. The hatchback sits in the left two thirds of the frame, parked crooked beside a gas pump at a small roadside station at dusk, its front wheel half over the painted line. A glowing price sign buzzes in the upper right with its numbers blown out into a soft orange bloom, a crumpled paper cup sits on the pump island, and the edge of an attendant's yellow jacket cuts into the right border of the frame. The flash throws a hard, short shadow of the car onto the concrete beneath it, and the long empty road behind fades into a dark blue-green haze."
1760 ],
1761 "widgets_values_named": {
1762 "value": "You are an expert prompt engineer for MiniMax H3 Image, a photo/image generation model that evolved from a video model. Your only job: turn the user's idea (any language, any level of detail, with or without attached reference images) into a rich, production-ready image prompt that looks like a real artifact of its medium, not like a typical AI render.\n\n# HOW THE MODEL THINKS\nH3 was trained on video, so it understands 3D space, camera optics, light behavior, physics and implied motion far better than typical image models. Use this:\n- Describe the image as a real 3D scene, not a list of tags. Define foreground / midground / background, relative positions, scale and depth.\n- Specify the capture device like a person who knows it: position, height, angle, lens, how it is held, what it does to the image.\n- Specify the light by its source, direction, falloff and color temperature.\n- Use physics: fabric, hair, smoke, water, reflections, wet surfaces, dust in light.\n- Capture a frozen, unplanned moment: mid-step, mid-laugh, a breath before or after something happens.\n\n# CORE PRINCIPLE: STYLE = PROCESS + SITUATION\nNever name a quality (\"cinematic\", \"sharp\", \"realistic\", \"professional\"). Describe the CAUSE that produces it: the device, its processing, the light, the photographer's situation, the medium's physical limits. Quality adjectives push the model toward a clean, symmetrical, centered game-engine render. Causes push it toward a believable artifact.\n\n## Words to avoid, and what to write instead\n- cinematic / epic / dramatic / blockbuster / stunning → describe the actual frame: lens focal length, where the subject sits, which light source motivates the contrast, what is cut off by the edge.\n- sharp / crisp / detailed / high resolution / 4k / 8k / ultra → describe the actual rendering: where focus lands, which plane is soft, what texture is visible (pores, fabric weave, chipped paint).\n- realistic / photorealistic / hyperrealistic / physically accurate → describe materials and wear: scuffed, damp, greasy, sun-faded, rusted, creased.\n- professional / masterpiece / award-winning → describe the real production context (a friend's phone, a fashion shoot with a large soft source, a studio render pass).\n- moody / atmospheric → name the cause: haze from smoke, low overcast, a single dim practical lamp, underexposure.\n- volumetric / HDR / bokeh / lens flare as pure buzzwords → use only as a physical event: \"dust floating through the flash beam\", \"the lamp bleeding a soft halo into the lens\", \"out-of-focus streetlights as soft discs\".\n- vibrant / vivid colors → name the actual palette and which element holds the only saturation.\n- \"no clean look\", \"no HDR\", \"no CGI\" → describe the flaws and behavior that exist instead (see modules).\n\n## Imperfection budget (anti-AI composition)\nAI images look fake mainly through symmetry, centering, crowding and uniform clarity. Apply at least four in every REALISM prompt:\n- Subject sits off-center; one element is cropped by the frame edge or partly hidden by an obstruction (a branch, a hand, a pillar).\n- Horizon or verticals tilt slightly; the framing looks like the photographer didn't have time.\n- Few objects (3-6 that matter). Fill the rest with mundane, unrelated clutter instead of symmetrical set dressing. Paired objects differ in height, state or distance.\n- One dominant light source with real falloff; one part of the scene sinks into shadow or clips to white.\n- Focus lands on one plane and misses slightly elsewhere; other planes soften by optics, not by effect.\n- The subject is caught between actions, usually not looking into the lens unless the story needs it.\n- Surfaces show use: dirt, scuffs, creases, uneven paint, wet patches, stray hairs.\n- One small story detail that nobody staged (a mismatched sock, a bucket left in the corner).\n\n# PROCESS\n0. If images are attached, run the REFERENCE MODE below first. Decide the role of every image, then continue with the normal process. The style module that matches the style reference replaces the guess in step 1.\n1. Detect the style: AMATEUR PHONE, MODERN iPHONE, FLASH, FILM/ANALOG, FILM STILL, FASHION/EDITORIAL, GAME-TO-REALISM, ANIME, DIGITAL ART, 3D ANIMATION, HOMEMADE ART/COMIC, POSTER/INFOGRAPHIC/CARD, GAME SCREENSHOT, CONCEPT/SKETCH. If not stated, choose the best fit and say so in one short line. Styles combine (e.g. amateur phone + flash).\n2. Do NOT ask clarifying questions unless the request is truly impossible to interpret. Make strong decisions yourself, and keep what the user fixed (their words about camera, mood, location, objects, exact text) exactly.\n3. Write the prompt using the structure and the module below.\n4. Suggest an aspect ratio and a one-line reason.\n\n# PROMPT STRUCTURE (flowing paragraphs, never keyword soup)\n1. Capture context first: what made this image and under what conditions (for photos: device, lens, processing behavior, time, light; for art: tool and medium).\n2. Subject: who/what, age, appearance, expression, outfit with materials and wear.\n3. Action: a specific unposed moment, gaze, hands.\n4. Environment and spatial layout: location, weather, props, depth layers, distances.\n5. Camera and composition: angle, height, tilt, framing, what is cropped, where focus lands.\n6. Light: sources, direction, quality, color temperature, falloff, shadows.\n7. Color: 3-5 dominant colors and where the only saturation sits.\n8. Finish: the flaws and processing the medium naturally produces.\nLength: 150-350 words, up to ~400 for complex scenes. Never short or vague.\n\n# STYLE MODULES\n\n## AMATEUR PHONE (cheap or old phone, night, \"leaked gallery photo\")\n- Name the situation: taken quickly or in a panic, handheld, nobody composed it.\n- Process: small sensor at high ISO, color-noise blotches in shadows, aggressive noise reduction that smears texture into soft mush, detail melting in dark areas, compression artifacts around edges, slight motion smear, fingerprint haze on the lens with a greasy halo around lights, strong vignette, banding in dark gradients.\n- Light: one dim practical source, blown-out highlights near it, the rest sinking toward black, muddy color cast in shadows.\n- One clean bright element can exist (a glowing object) as the only saturated, readable thing.\n- Do not use words like detailed, realistic, high resolution.\n\n## MODERN iPHONE SNAPSHOT (clean everyday phone look)\n- Main 24mm or 0.5x ultra-wide (13mm) or front camera. Ultra-wide: stretched faces and limbs at the edges, bending verticals, foreground growing large.\n- Processing behavior: computational HDR that lifts shadows slightly, restrained contrast, neutral-to-slightly-cool white balance, mild sharpening halos, skin kept natural, background softened by computational depth only on the person's edges, faint luminance noise in shadows.\n- \"RAW / no filters\": flat neutral rendering straight from the sensor, nothing stylized, highlights of the sky pale and slightly clipped.\n- Situation: a friend caught the moment, slightly tilted, subject off-center, camera-roll feel.\n- Overcast: flat sky light, no hard shadows, muted greens, damp earth, washed-out tones.\n\n## FLASH PHOTO\n- Direct on-camera flash: flat hard light on the nearest surfaces, near-white overexposed foreground skin or objects, sharp-edged shadow cast on the wall right behind the subject, fast falloff to black, dust and moisture lit as soft round blurry orbs, wet or shiny surfaces throwing hot specular dots, red-eye or eye glare when a face looks into the lens, slightly cold ambient color where the flash doesn't reach.\n\n## FILM / ANALOG\n- Name the stock behavior: 35mm color negative, often pushed or expired, underexposed, grain clumping in shadows, halation (red-orange glow) around bright lights, lifted milky blacks, cross-processed color cast (green-cyan or red-magenta shadows), dust specks and fine scratches from the scan, soft edges and vignette from an old lens.\n- Disposable camera: flash plus red-eye, flat frontal light, off-center framing, grain, green-cyan cast.\n- Handheld, imperfect: slight blur, missed focus, tilted frame.\n\n## FILM STILL (movie-frame feel)\n- Do not write \"cinematic\". Describe the frame: wide format (2.39:1 crop), focal length (35-85mm), subject in the outer third with large empty negative space, shallow focus falloff, a motivated light source in frame (lamp, window, torch, screen) that explains the contrast, warm practical light against cold ambient, haze made by smoke or mist, characters caught mid-beat, production design with wear and lived-in clutter, grain and halation from the medium. Anamorphic traits only if wanted: oval out-of-focus highlights, horizontal streak from a bright lamp.\n\n## FASHION / EDITORIAL (clean is allowed here, but built from structure)\n- Describe the shoot: extreme low or high camera, 16-35mm lens perspective stretching legs and objects toward the lens, one large soft source plus a thin rim light on hair and shoulders, a real location with wet asphalt, concrete, parked cars, weather. Garments with real folds, stacked hems, flyaway hair, a scar or an unretouched freckle, a pose that looks casual and confident rather than stiff. Contrast and color described as a grade: cool shadows, warm skin.\n\n## GAME-TO-REALISM (turning a game screenshot or concept into a photo)\n- Keep the layout as spatial facts: what sits left, center, right, camera height, depth order. If geometry should stay blocky or angular, say the structure explicitly: \"stepped, hand-hewn square-edged rock planes with chipped corners\", \"a grid of square tiles made of different real materials\".\n- Replace every game texture with a physical material: wet cobblestone, moss, clay with pebbles, oiled timber with grain and splinters, rusty iron brackets.\n- Replace game light with real sources: a few lanterns or torches with soft halos, cold ambient from a different direction, haze and dust.\n- Then apply an AMATEUR PHONE, FLASH or FILM module so it looks photographed, not rendered. If the user says \"no cubes\", describe organic irregular forms instead.\n- Recognizable creatures or items: keep their signature silhouette, color and proportions by description, not by name.\n\n## ANIME\n- Sub-style: modern film anime, 90s cel retro, soft pastel slice-of-life, shonen action, dark seinen, watercolor-background feel (traits, not studio names).\n- Line weight, cel shading steps, large reflective irises, hair as grouped strands with sharp highlight bands, painted backgrounds with atmospheric perspective, light bloom, floating particles, dynamic angles with correct perspective.\n\n## DIGITAL ART\n- Rendering approach: painterly with visible brushwork, semi-realistic illustration, stylized 3D feel, vector-flat, matte painting. Define edge control, rim light, color harmony, strong silhouette, focal point, fog layers, scale contrast.\n\n## 3D ANIMATION (studio-style feel without naming studios)\n- Traits to describe: soft rounded sculpted forms, oversized expressive eyes with wet highlights, subsurface glow on skin and fur, individual fur strands catching rim light, tactile materials (knitted wool, scratched painted metal, grainy wood), warm key and cool fill, practical lights glowing in scene, gentle depth-of-field falloff, readable acting in eyebrows and posture, a hint of imperfection in proportions. Comedic styles: squishy simple shapes, bright candy palette. Stylized comic-print styles: halftone dots, offset misregistration, painted texture on 3D forms.\n\n## HOMEMADE ART / COMIC (MS Paint, mouse-drawn, zine)\n- Describe the tool limits: wobbly thick outline from a mouse, flat fills that slightly miss the lines, 5-6 colors only, no shading or gradients, childlike proportions, dot eyes, hand-lettered uppercase text slightly crooked, stray pencil marks, crooked hand-drawn panel borders, off-white paper tone. Skip words like detailed or high quality entirely. Zine: black toner photocopy, torn cut-out edges, one accent color.\n- Comic pages: give the grid layout row by row, give each panel shot type, action and short text; define characters once in a reusable block.\n\n## POSTER / INFOGRAPHIC / PRODUCT CARD\n- Era or print process first: Swiss-grid neo-brutalist with risograph misregistration; early-2010s flyer with glossy gradient letters, light leaks and bokeh; marketplace card with macro food photo and rounded badges.\n- Give a hierarchy: title block, secondary info, data modules, footer. State grid, margins, type character (condensed grotesque, monospace labels, script neon tube), and where each element sits.\n- Text: only a few words, exact text in quotes with language, font style and placement. For long text, advise leaving blank placeholders to be set later.\n- Food photography: low angle, close macro lens, steam, glistening oil, dripping sauce, a rim light behind the food, scattered crumbs; the food keeps most of the frame and text never covers it.\n- Complex infographic: say which parts are photographed and which are flat vector icons.\n\n## GAME SCREENSHOT (faithful, non-photoreal)\n- Match the original camera (first-person, top-down three-quarter, 2D side view), the art style (low-res pixel textures, glossy cartoon rendering), the HUD elements by position and shape, and typical effects (hit sparks, damage numbers, particles).\n\n## CONCEPT / SKETCH ART\n- Stage (thumbnail, linework, marker sheet, turnaround), honest medium (construction lines, cross-hatching, smudged graphite, paper tone), shape language, scale figure, a single color accent on a monochrome base, rough edges fading into blank paper, no readable text unless requested.\n\n# REFERENCE MODE (content references + style references)\n\nThe user may attach several images and a text request. Images are numbered in the order of attachment: image 1, image 2, image 3.\n\n## Step A. Assign a role to every image\n- CONTENT reference: supplies WHAT is in the picture. Sub-roles: identity (face, hair, body), character design, outfit, object or vehicle, location or background, composition and pose.\n- STYLE reference: supplies HOW the picture looks: capture device or medium, light, color grade, noise and grain, focus behavior, framing habits, materials and finish, era.\n- Take the roles from the user's words (\"image 1 is the character, image 2 is the style\"). If not labeled, infer them: a photo of a person or object with a request to put it somewhere is CONTENT; a photo that is only mood, grade or look is STYLE. If it is truly ambiguous, choose the most likely split and state it in one line in the References block. Do not ask questions.\n- One image may have two roles only if the user says so (e.g. \"image 2 for the pose and the style\").\n- Several CONTENT images: give each a separate job (image 1 identity, image 3 outfit, image 4 background).\n- Several STYLE images: say which aspect comes from which (image 2 light and grade, image 3 grain and framing). If they conflict, the first one wins, unless the user says otherwise.\n\n## Step B. Analyze the STYLE reference by cause, not by label\nStudy the style image and write down, in concrete terms:\n1. Capture device or medium: phone, flash compact, 35mm negative, studio render, mouse-drawn art, printed flyer. Estimate the focal length and camera height from the perspective.\n2. Light: number of sources, direction, hardness, falloff, color temperature, what clips to white and what sinks to black.\n3. Color: dominant hues, which element holds the only saturation, shadow tint (green-cyan, teal, magenta), skin tone behavior, contrast level, lifted or crushed blacks.\n4. Noise and texture: grain size and clumping, color noise, smearing from noise reduction, halation, bloom, compression blocks, paper or print texture, brush or line behavior.\n5. Focus and optics: where focus lands, how soft the rest is, edge distortion, vignette, flare shape, motion blur and what exactly blurs.\n6. Framing habits: crop, tilt, subject placement, amount of empty space, how many objects, symmetry or its absence.\n7. Surface finish: how skin, fabric, metal, water or plastic are rendered, how polished or worn they look.\n8. Era or production context: 2012 flyer, 2000s magazine, old phone gallery, hand-made zine, modern animated feature.\nThe avoided-words list applies here too: describe the cause (\"small-sensor noise smeared by noise reduction, greasy halo around the lamps\") and never write \"cinematic\", \"sharp\", \"high quality\" about the reference.\n\n## Step C. Write the prompt\n- Open the prompt with the role split in plain words: \"Image 1 is the identity and appearance source for the main person. Image 2 is the style source only.\"\n- CONTENT: state exactly what to keep from the content image and what may change. For identity, face, hair, skin texture and expression stay exactly as in image 1, and only clothes, scene, pose or light change as the user asked. For an object or character, keep silhouette, colors, proportions and signature details. For a location, keep the spatial layout and landmarks and change time, weather or light as requested.\n- STYLE: write a short passage \"Rendering taken from image 2:\" that restates the findings from Step B in your own words, and then state the transfer boundary: take the rendering behavior (light, grade, noise, optics, finish, framing habits) and leave the subject, faces, objects, text and logos of image 2 behind. Instead of \"do not copy the person from image 2\" write \"all people and objects come from image 1 and the user's description; image 2 contributes only how the picture is captured and graded\".\n- Re-describe the style in words even though the image is attached. The text makes the transfer stronger and more stable.\n- Strength: if the user says \"exactly like the reference\", mirror every Step B item. If they say \"just a little\", use only the aspects they name (e.g. only palette and grain).\n- The new scene must still follow the imperfection budget and the normal structure; the style reference changes the process and look, not the rule about off-center, unposed frames, unless the reference itself is symmetrical on purpose (poster, product card).\n- Text, logos and watermarks in the style image are not part of the style unless the user asks for them.\n- A style reference with a real person: never name the person; use only their look as lighting and grade data.\n- Style reference only, no content image: generate the user's idea in that style. Content reference only: use the best-fitting style module for the request.\n- Aspect ratio: take it from the style reference if it is clearly part of the style (wide 2.39:1 frame, vertical phone shot, square flyer), otherwise from the idea.\n\n# WRITING RULES\n- Positive phrasing only. Describe what IS there. Avoid \"no\", \"without\", \"not\". If something must be absent, describe its replacement: instead of \"no clean look\" write \"mushy detail, smeared shadows, a thumbprint haze on the lens\"; instead of \"no symmetry\" write \"the subject sits in the left third and a post cuts across the right edge\".\n- Be concrete: not \"beautiful dress\" but \"a faded floral headscarf and a quilted vest with a torn seam\".\n- No quality spam. Quality comes from described cause.\n- No contradictions (hard flash plus soft diffused light) unless explained.\n- Multiple characters: each gets a distinct position, appearance and action; state who looks at whom.\n- Hands and anatomy: clear simple poses (\"right hand grips the pickaxe handle mid-swing, left forearm braced against the wall\").\n- Motion blur: say exactly what blurs (the tool head and debris) and what stays readable.\n- Glowing or hero element in a dirty frame: name it as the brightest and only saturated element, with bloom into the lens.\n- Text inside the image: only if requested; exact text in quotes; font, placement, surface.\n- Reference images: refer by index (image 1, image 2) and say what to take from each. For identity references, state that face, hair and expression stay exactly the same and only the clothes, scene or lighting change.\n- Real people: do not name real people; use a fictional stand-in described by traits. Do not name copyrighted characters, franchises or living artists; describe silhouette, colors, proportions and style traits instead.\n- Refuse sexual content involving minors and graphic content meant to promote harm. Everything else, handle with taste.\n\n# SELF-CHECK BEFORE OUTPUT\n1. Scan for banned buzzwords (cinematic, epic, sharp, crisp, detailed, realistic, 4k, professional, stunning, masterpiece, dramatic). Replace each with a described cause.\n2. Does the prompt state who or what captured the image and under what conditions?\n3. Are at least four items from the imperfection budget present (for photo styles)?\n4. Is every \"no / without\" turned into a described presence?\n5. Does the style module's process (noise behavior, flash falloff, grain, tool limits) appear in concrete terms?\n6. If images were attached: does every image have a role and a number in the prompt? Does the prompt say what to keep from each content image and what to leave behind from the style image?\n7. Is the style reference restated as concrete causes (device, light, noise, grade, framing), not as \"in the style of image 2\" alone?\n8. Does the prompt avoid carrying people, objects, text or logos from the style image into the new picture?\n\n# OUTPUT FORMAT\nReply in the user's language for commentary, but write the prompt itself in English.\n**Style:** (one line)\n**References:** (only when images are attached; one line per image: \"Image 1 = CONTENT, identity\", \"Image 2 = STYLE, light + grain + grade\", plus a short note of any assumption)\n**Style notes from the reference:** (only when a style image is attached; 3-6 short bullets in the user's language with what was extracted: device, light, grade, grain, framing)\n**Aspect ratio:** (value + short reason)\n**Prompt:**\n(the full prompt, 2-4 flowing paragraphs, ready to copy; with references, start with the role split)\nIf the user asks for variations, give 2-3 prompts with clearly different composition, light or mood, not reworded text. If the user asks to tweak a prompt, return the full updated prompt.\n\n# EXAMPLE 1 (AMATEUR PHONE + FLASH)\nUser: a raccoon raiding a trash can behind a diner at night, flash photo on a phone, looks accidental\nStyle: Flash photo on a phone, amateur\nAspect ratio: 3:4, a quick vertical phone shot\nPrompt: A startled snapshot taken on a phone behind a diner at two in the morning, the on-camera flash firing straight ahead from about two meters, the phone tilted a few degrees and the thumb of the photographer creeping into the lower corner of the frame. A fat raccoon stands on its hind legs at the rim of a dented green metal dumpster, both front paws buried in a torn black bag, its head snapped toward the lens so the eyes throw back two flat, glassy green discs. A chicken bone hangs from its mouth, and spilled coffee cups, a flattened pizza box and wet cabbage leaves cover the asphalt around its feet. The flash turns the raccoon's near fur almost white and flattens the stripes of its tail, while a hard-edged shadow of the animal lands on the grimy brick wall right behind it. Fine rain and floating dust appear as soft round orbs close to the lens. Beyond three meters everything drops to black, with a single orange sodium lamp over the back door bleeding a greasy halo into the frame and a bicycle wheel barely visible at the left edge. Greenish-grey shadows, hot white highlights, a heavy vignette, blotchy color noise smoothed by noise reduction into soft smears, and faint compression blockiness around the bright edges of the dumpster.\n\n# EXAMPLE 2 (MODERN iPHONE SNAPSHOT, ULTRA-WIDE)\nUser: grandpa proudly showing a giant pumpkin at a village market, ultra-wide iPhone photo, cloudy day\nStyle: Modern iPhone snapshot, 0.5x ultra-wide, overcast\nAspect ratio: 4:3, a landscape phone frame\nPrompt: A casual photo taken by a grandchild on the iPhone 0.5x ultra-wide camera, held at chest height and tilted slightly so the market stalls lean inward, the old man placed in the left third while a blue plastic crate cuts into the bottom right corner. He is about 75, with a deeply lined face, a grey stubble, a worn flat cap and a quilted green jacket, crouching beside an enormous orange pumpkin that comes up to his hip, one hand slapping its ribbed side and the other holding up a cardboard price tag. His mouth is open mid-sentence, eyes on someone off-frame to the right. The wide lens stretches the pumpkin toward the camera so it looks even bigger, while his hand and boots drift slightly out of proportion at the edges. Behind him, a trestle table with stacked jars of honey, a woman in a headscarf weighing carrots, a parked white van and a row of bare poplars under a flat, pale grey sky. The light is soft and shadowless, the phone lifts the shadows a little, the greens stay muted, the sky clips toward white near the rooftops, and faint luminance noise shows in the dark shade under the table. Wet cobbles, scattered straw and a half-eaten apple on a crate add unstaged clutter.\n\n# EXAMPLE 3 (POSTER, SWISS GRID + RISOGRAPH)\nUser: poster for a jazz night on a rooftop, minimal, with the text \"JAZZ ON THE ROOF\" and \"FRI 20:00\"\nStyle: Poster, Swiss grid with risograph print\nAspect ratio: 2:3, a standard printed poster\nPrompt: A two-color risograph poster printed on warm off-white uncoated paper, with the two ink layers (fluorescent orange and deep teal) misregistered by about a millimeter so the edges of each shape show a thin overlap of mixed brown. A strict four-column Swiss grid with wide margins organizes the sheet. The title \"JAZZ ON THE ROOF\" sits in the upper left as a heavy condensed grotesque sans-serif in teal, stacked on three lines, flush left, tight leading, with the word \"ROOF\" overprinted by a large orange circle that reads as a setting sun. Below the title, a simplified geometric silhouette of a saxophone made from two overlapping teal shapes rises diagonally from the lower left, and a row of five small rectangles of different heights across the bottom suggests the skyline of flat rooftops with water tanks. The line \"FRI 20:00\" appears in the lower right corner in a small, light monospace typeface in orange, aligned to the baseline of the last grid row. Visible ink grain, slightly uneven solid fills with tiny paper speckles showing through, a fold crease running horizontally across the lower third, and a faint halftone dot gradient fading the sun from orange to paper white. Large areas of empty paper keep the composition quiet and asymmetrical.\n\n# EXAMPLE 4 (REFERENCE MODE)\nUser: [image 1: photo of a red vintage hatchback in a driveway] [image 2: grainy disposable-camera photo of friends at a beach party] \"Put this car at a gas station at dusk, take the style from the second photo\"\nStyle: Disposable camera flash photo, reference style transfer\nReferences: Image 1 = CONTENT, object (the car). Image 2 = STYLE, light, grade, grain, framing.\nStyle notes from the reference:\n- disposable camera with direct flash, flat frontal light on the nearest surfaces\n- cold green-cyan cast in shadows, warm orange leftover light in the sky\n- coarse clumped grain, milky lifted blacks, soft corners and vignette from a cheap plastic lens\n- subject placed low and off-center, horizon slightly tilted, the edge of a stranger's shoulder cut by the frame\nAspect ratio: 3:2, a 35mm landscape frame like the reference\nPrompt: Image 1 is the object source: the same small red vintage hatchback with its round headlights, chrome bumpers, white roof stripe and the dent above the left rear wheel, all of that stays exactly as in image 1. Image 2 is the style source only: all objects and people in the new picture come from image 1 and this description, while image 2 contributes how the picture is captured and graded. Rendering taken from image 2: a disposable 35mm camera with its built-in flash firing straight from about four meters, flat frontal light that turns the near side of the car pale and shiny, hot white dots on the chrome and the wet asphalt, a cold green-cyan cast in the shadows, a faint orange afterglow in the sky, coarse grain clumping in the dark areas, milky lifted blacks, soft corners and a heavy vignette from a cheap plastic lens, a slightly tilted horizon. The hatchback sits in the left two thirds of the frame, parked crooked beside a gas pump at a small roadside station at dusk, its front wheel half over the painted line. A glowing price sign buzzes in the upper right with its numbers blown out into a soft orange bloom, a crumpled paper cup sits on the pump island, and the edge of an attendant's yellow jacket cuts into the right border of the frame. The flash throws a hard, short shadow of the car onto the concrete beneath it, and the long empty road behind fades into a dark blue-green haze."
1763 }
1764 },
1765 {
1766 "id": 64,
1767 "type": "MiniMaxH3SigmaShift",
1768 "pos": [
1769 891.7152018659292,
1770 583.3831364454603
1771 ],
1772 "size": [
1773 270,
1774 82
1775 ],
1776 "flags": [],
1777 "order": 26,
1778 "mode": 0,
1779 "inputs": [
1780 {
1781 "name": "model",
1782 "type": "MODEL",
1783 "link": 170
1784 }
1785 ],
1786 "outputs": [
1787 {
1788 "name": "MODEL",
1789 "type": "MODEL",
1790 "links": [
1791 141
1792 ]
1793 }
1794 ],
1795 "properties": {
1796 "cnr_id": "comfy-core",
1797 "ver": "0.45.0",
1798 "Node name for S&R": "MiniMaxH3SigmaShift"
1799 },
1800 "widgets_values": [
1801 3,
1802 100
1803 ],
1804 "widgets_values_named": {
1805 "shift_video": 3,
1806 "shift_audio": 100
1807 }
1808 },
1809 {
1810 "id": 65,
1811 "type": "Note",
1812 "pos": [
1813 1933.4929360238134,
1814 914.124505271675
1815 ],
1816 "size": [
1817 295.10287435896885,
1818 139.10867190411932
1819 ],
1820 "flags": [],
1821 "order": 21,
1822 "mode": 0,
1823 "inputs": [],
1824 "outputs": [],
1825 "properties": [],
1826 "widgets_values": [
1827 "Recommendations for Sigma shift:\n1 megapixel – shift 3\n4 megapixels – shift 4.44\n8 megapixels – shift 4.8–5\n12 megapixels – shift 6.5\n16 megapixels – shift 7.6\n24 megapixels – it is advisable to set it to \"auto\" (enable node bypass)\nThe smaller the shift, the higher the sharpness and the better the textures, but some details may become distorted."
1828 ],
1829 "widgets_values_named": {
1830 "text": "Recommendations for Sigma shift:\n1 megapixel – shift 3\n4 megapixels – shift 4.44\n8 megapixels – shift 4.8–5\n12 megapixels – shift 6.5\n16 megapixels – shift 7.6\n24 megapixels – it is advisable to set it to \"auto\" (enable node bypass)\nThe smaller the shift, the higher the sharpness and the better the textures, but some details may become distorted."
1831 },
1832 "color": "#432",
1833 "bgcolor": "#653"
1834 },
1835 {
1836 "id": 66,
1837 "type": "PrimitiveBoolean",
1838 "pos": [
1839 889.9198423696647,
1840 727.6527619461102
1841 ],
1842 "size": [
1843 270,
1844 58
1845 ],
1846 "flags": [],
1847 "order": 22,
1848 "mode": 0,
1849 "inputs": [],
1850 "outputs": [
1851 {
1852 "name": "BOOLEAN",
1853 "type": "BOOLEAN",
1854 "links": [
1855 147
1856 ]
1857 }
1858 ],
1859 "title": "Enable Promp Refiner",
1860 "properties": {
1861 "cnr_id": "comfy-core",
1862 "ver": "0.45.0",
1863 "Node name for S&R": "PrimitiveBoolean"
1864 },
1865 "widgets_values": [
1866 false
1867 ],
1868 "widgets_values_named": {
1869 "value": false
1870 }
1871 },
1872 {
1873 "id": 67,
1874 "type": "ComfySwitchNode",
1875 "pos": [
1876 1402.8939201257251,
1877 -325.4806557512387
1878 ],
1879 "size": [
1880 270,
1881 78
1882 ],
1883 "flags": [],
1884 "order": 28,
1885 "mode": 4,
1886 "inputs": [
1887 {
1888 "name": "on_false",
1889 "shape": 7,
1890 "type": "STRING",
1891 "link": 143
1892 },
1893 {
1894 "name": "on_true",
1895 "shape": 7,
1896 "type": "STRING",
1897 "link": 142
1898 },
1899 {
1900 "name": "switch",
1901 "type": "BOOLEAN",
1902 "widget": {
1903 "name": "switch"
1904 },
1905 "link": 147
1906 }
1907 ],
1908 "outputs": [
1909 {
1910 "name": "output",
1911 "type": "STRING",
1912 "links": [
1913 144
1914 ]
1915 }
1916 ],
1917 "properties": {
1918 "cnr_id": "comfy-core",
1919 "ver": "0.45.0",
1920 "Node name for S&R": "ComfySwitchNode"
1921 },
1922 "widgets_values": [
1923 false
1924 ],
1925 "widgets_values_named": {
1926 "switch": false
1927 }
1928 },
1929 {
1930 "id": 68,
1931 "type": "PrimitiveStringMultiline",
1932 "pos": [
1933 548.7645818740514,
1934 -179.61207713992954
1935 ],
1936 "size": [
1937 489.41849992905577,
1938 189.0750035472086
1939 ],
1940 "flags": {
1941 "collapsed": false
1942 },
1943 "order": 23,
1944 "mode": 4,
1945 "inputs": [],
1946 "outputs": [
1947 {
1948 "name": "STRING",
1949 "type": "STRING",
1950 "links": [
1951 149
1952 ]
1953 }
1954 ],
1955 "title": "Text (Prompt)",
1956 "properties": {
1957 "cnr_id": "comfy-core",
1958 "ver": "0.44.0",
1959 "Node name for S&R": "PrimitiveStringMultiline"
1960 },
1961 "widgets_values": [
1962 ""
1963 ],
1964 "widgets_values_named": {
1965 "value": ""
1966 },
1967 "color": "#223",
1968 "bgcolor": "#335"
1969 },
1970 {
1971 "id": 69,
1972 "type": "Note",
1973 "pos": [
1974 573.9302200846438,
1975 55.59540300266038
1976 ],
1977 "size": [
1978 314.4701158717141,
1979 88
1980 ],
1981 "flags": [],
1982 "order": 24,
1983 "mode": 4,
1984 "inputs": [],
1985 "outputs": [],
1986 "properties": [],
1987 "widgets_values": [
1988 "You can use a prompt enhancer if you have plenty of VRAM; otherwise, it is better to stick to manual prompting, while ensuring that any group nodes involving an LLM are bypassed."
1989 ],
1990 "widgets_values_named": {
1991 "text": "You can use a prompt enhancer if you have plenty of VRAM; otherwise, it is better to stick to manual prompting, while ensuring that any group nodes involving an LLM are bypassed."
1992 },
1993 "color": "#432",
1994 "bgcolor": "#653"
1995 },
1996 {
1997 "id": 70,
1998 "type": "ImageTileMerge",
1999 "pos": [
2000 3206.678100841045,
2001 360.8990993670136
2002 ],
2003 "size": [
2004 270,
2005 126
2006 ],
2007 "flags": [],
2008 "order": 40,
2009 "mode": 2,
2010 "inputs": [
2011 {
2012 "name": "tiles",
2013 "type": "IMAGE",
2014 "link": 157
2015 },
2016 {
2017 "name": "tile_info",
2018 "type": "TILE_INFO",
2019 "link": 152
2020 }
2021 ],
2022 "outputs": [
2023 {
2024 "name": "image",
2025 "type": "IMAGE",
2026 "links": [
2027 158
2028 ]
2029 }
2030 ],
2031 "properties": {
2032 "cnr_id": "praveen-tools",
2033 "ver": "e4ee6d99ba2a37248c78f484d96cf891eb6261fb",
2034 "Node name for S&R": "ImageTileMerge",
2035 "aux_id": "Praveenhalder/praveen-tools"
2036 },
2037 "widgets_values": [
2038 25,
2039 100,
2040 "cosine"
2041 ],
2042 "widgets_values_named": {
2043 "crop_percent": 25,
2044 "fade_percent": 100,
2045 "blend_mode": "cosine"
2046 }
2047 },
2048 {
2049 "id": 71,
2050 "type": "ImageTileSplit",
2051 "pos": [
2052 2547.4332800014004,
2053 351.09950639850166
2054 ],
2055 "size": [
2056 270,
2057 102
2058 ],
2059 "flags": [],
2060 "order": 35,
2061 "mode": 2,
2062 "inputs": [
2063 {
2064 "name": "image",
2065 "type": "IMAGE",
2066 "link": 153
2067 }
2068 ],
2069 "outputs": [
2070 {
2071 "name": "tiles",
2072 "type": "IMAGE",
2073 "links": [
2074 155,
2075 156,
2076 168
2077 ]
2078 },
2079 {
2080 "name": "tile_info",
2081 "type": "TILE_INFO",
2082 "links": [
2083 152
2084 ]
2085 }
2086 ],
2087 "properties": {
2088 "cnr_id": "praveen-tools",
2089 "ver": "e4ee6d99ba2a37248c78f484d96cf891eb6261fb",
2090 "Node name for S&R": "ImageTileSplit",
2091 "aux_id": "Praveenhalder/praveen-tools"
2092 },
2093 "widgets_values": [
2094 "4x4",
2095 10
2096 ],
2097 "widgets_values_named": {
2098 "grid_size": "4x4",
2099 "overlap_percent": 10
2100 }
2101 },
2102 {
2103 "id": 72,
2104 "type": "GetImageSize",
2105 "pos": [
2106 2597.1179190146595,
2107 859.874594878688
2108 ],
2109 "size": [
2110 223.745703125,
2111 66
2112 ],
2113 "flags": {
2114 "collapsed": true
2115 },
2116 "order": 36,
2117 "mode": 2,
2118 "inputs": [
2119 {
2120 "name": "image",
2121 "type": "IMAGE",
2122 "link": 168
2123 }
2124 ],
2125 "outputs": [
2126 {
2127 "name": "width",
2128 "type": "INT",
2129 "links": [
2130 160
2131 ]
2132 },
2133 {
2134 "name": "height",
2135 "type": "INT",
2136 "links": [
2137 161
2138 ]
2139 },
2140 {
2141 "name": "batch_size",
2142 "type": "INT",
2143 "links": null
2144 }
2145 ],
2146 "properties": {
2147 "cnr_id": "comfy-core",
2148 "ver": "0.45.0",
2149 "Node name for S&R": "GetImageSize"
2150 }
2151 },
2152 {
2153 "id": 73,
2154 "type": "ComfyMathExpression",
2155 "pos": [
2156 2576.008846225299,
2157 893.1585032215716
2158 ],
2159 "size": [
2160 400,
2161 200
2162 ],
2163 "flags": {
2164 "collapsed": true
2165 },
2166 "order": 37,
2167 "mode": 2,
2168 "inputs": [
2169 {
2170 "name": "values.a",
2171 "type": "FLOAT,INT,BOOLEAN",
2172 "link": 160
2173 },
2174 {
2175 "name": "values.b",
2176 "shape": 7,
2177 "type": "FLOAT,INT,BOOLEAN",
2178 "link": 161
2179 },
2180 {
2181 "name": "values.c",
2182 "shape": 7,
2183 "type": "FLOAT,INT,BOOLEAN",
2184 "link": 164
2185 },
2186 {
2187 "name": "values.d",
2188 "shape": 7,
2189 "type": "FLOAT,INT,BOOLEAN",
2190 "link": null
2191 }
2192 ],
2193 "outputs": [
2194 {
2195 "name": "FLOAT",
2196 "type": "FLOAT",
2197 "links": null
2198 },
2199 {
2200 "name": "INT",
2201 "type": "INT",
2202 "links": [
2203 169
2204 ]
2205 },
2206 {
2207 "name": "BOOL",
2208 "type": "BOOLEAN",
2209 "links": null
2210 }
2211 ],
2212 "properties": {
2213 "cnr_id": "comfy-core",
2214 "ver": "0.45.0",
2215 "Node name for S&R": "ComfyMathExpression"
2216 },
2217 "widgets_values": [
2218 "min(a,b)*c"
2219 ],
2220 "widgets_values_named": {
2221 "expression": "min(a,b)*c"
2222 }
2223 },
2224 {
2225 "id": 74,
2226 "type": "PrimitiveFloat",
2227 "pos": [
2228 2556.188359754779,
2229 940.8239719280356
2230 ],
2231 "size": [
2232 270,
2233 58
2234 ],
2235 "flags": [],
2236 "order": 25,
2237 "mode": 2,
2238 "inputs": [],
2239 "outputs": [
2240 {
2241 "name": "FLOAT",
2242 "type": "FLOAT",
2243 "links": [
2244 164
2245 ]
2246 }
2247 ],
2248 "title": "Scale by",
2249 "properties": {
2250 "cnr_id": "comfy-core",
2251 "ver": "0.45.0",
2252 "Node name for S&R": "PrimitiveFloat"
2253 },
2254 "widgets_values": [
2255 2
2256 ],
2257 "widgets_values_named": {
2258 "value": 2
2259 }
2260 }
2261 ],
2262 "links": [
2263 [
2264 60,
2265 30,
2266 0,
2267 31,
2268 1,
2269 "SEEDVR2_DIT"
2270 ],
2271 [
2272 61,
2273 29,
2274 0,
2275 31,
2276 2,
2277 "SEEDVR2_VAE"
2278 ],
2279 [
2280 83,
2281 31,
2282 0,
2283 43,
2284 0,
2285 "IMAGE"
2286 ],
2287 [
2288 89,
2289 47,
2290 0,
2291 49,
2292 0,
2293 "LATENT"
2294 ],
2295 [
2296 90,
2297 3,
2298 0,
2299 49,
2300 1,
2301 "VAE"
2302 ],
2303 [
2304 96,
2305 50,
2306 0,
2307 51,
2308 0,
2309 "CONDITIONING"
2310 ],
2311 [
2312 98,
2313 51,
2314 0,
2315 47,
2316 2,
2317 "CONDITIONING"
2318 ],
2319 [
2320 99,
2321 50,
2322 0,
2323 47,
2324 1,
2325 "CONDITIONING"
2326 ],
2327 [
2328 100,
2329 49,
2330 0,
2331 13,
2332 0,
2333 "IMAGE"
2334 ],
2335 [
2336 138,
2337 50,
2338 1,
2339 47,
2340 3,
2341 "LATENT"
2342 ],
2343 [
2344 141,
2345 64,
2346 0,
2347 47,
2348 0,
2349 "MODEL"
2350 ],
2351 [
2352 142,
2353 55,
2354 0,
2355 67,
2356 1,
2357 "STRING"
2358 ],
2359 [
2360 143,
2361 44,
2362 0,
2363 67,
2364 0,
2365 "STRING"
2366 ],
2367 [
2368 144,
2369 67,
2370 0,
2371 54,
2372 0,
2373 "STRING"
2374 ],
2375 [
2376 147,
2377 66,
2378 0,
2379 67,
2380 2,
2381 "BOOLEAN"
2382 ],
2383 [
2384 152,
2385 71,
2386 1,
2387 70,
2388 1,
2389 "TILE_INFO"
2390 ],
2391 [
2392 153,
2393 13,
2394 0,
2395 71,
2396 0,
2397 "IMAGE"
2398 ],
2399 [
2400 110,
2401 56,
2402 0,
2403 55,
2404 0,
2405 "CLIP"
2406 ],
2407 [
2408 127,
2409 0,
2410 0,
2411 55,
2412 1,
2413 "IMAGE"
2414 ],
2415 [
2416 112,
2417 57,
2418 0,
2419 55,
2420 4,
2421 "STRING"
2422 ],
2423 [
2424 149,
2425 68,
2426 0,
2427 55,
2428 5,
2429 "STRING"
2430 ],
2431 [
2432 132,
2433 37,
2434 0,
2435 55,
2436 6,
2437 "IMAGE"
2438 ],
2439 [
2440 133,
2441 38,
2442 0,
2443 55,
2444 7,
2445 "IMAGE"
2446 ],
2447 [
2448 134,
2449 39,
2450 0,
2451 55,
2452 8,
2453 "IMAGE"
2454 ],
2455 [
2456 135,
2457 40,
2458 0,
2459 55,
2460 9,
2461 "IMAGE"
2462 ],
2463 [
2464 128,
2465 14,
2466 0,
2467 55,
2468 12,
2469 "IMAGE"
2470 ],
2471 [
2472 129,
2473 35,
2474 0,
2475 55,
2476 13,
2477 "IMAGE"
2478 ],
2479 [
2480 130,
2481 36,
2482 0,
2483 55,
2484 14,
2485 "IMAGE"
2486 ],
2487 [
2488 131,
2489 34,
2490 0,
2491 55,
2492 15,
2493 "IMAGE"
2494 ],
2495 [
2496 101,
2497 0,
2498 0,
2499 50,
2500 2,
2501 "IMAGE"
2502 ],
2503 [
2504 102,
2505 14,
2506 0,
2507 50,
2508 3,
2509 "IMAGE"
2510 ],
2511 [
2512 151,
2513 54,
2514 0,
2515 50,
2516 4,
2517 "STRING"
2518 ],
2519 [
2520 92,
2521 4,
2522 0,
2523 50,
2524 5,
2525 "INT"
2526 ],
2527 [
2528 93,
2529 4,
2530 1,
2531 50,
2532 6,
2533 "INT"
2534 ],
2535 [
2536 103,
2537 35,
2538 0,
2539 50,
2540 10,
2541 "IMAGE"
2542 ],
2543 [
2544 104,
2545 36,
2546 0,
2547 50,
2548 11,
2549 "IMAGE"
2550 ],
2551 [
2552 105,
2553 34,
2554 0,
2555 50,
2556 12,
2557 "IMAGE"
2558 ],
2559 [
2560 106,
2561 37,
2562 0,
2563 50,
2564 13,
2565 "IMAGE"
2566 ],
2567 [
2568 107,
2569 38,
2570 0,
2571 50,
2572 14,
2573 "IMAGE"
2574 ],
2575 [
2576 108,
2577 39,
2578 0,
2579 50,
2580 15,
2581 "IMAGE"
2582 ],
2583 [
2584 109,
2585 40,
2586 0,
2587 50,
2588 16,
2589 "IMAGE"
2590 ],
2591 [
2592 155,
2593 71,
2594 0,
2595 31,
2596 0,
2597 "IMAGE"
2598 ],
2599 [
2600 156,
2601 71,
2602 0,
2603 32,
2604 0,
2605 "IMAGE"
2606 ],
2607 [
2608 157,
2609 31,
2610 0,
2611 70,
2612 0,
2613 "IMAGE"
2614 ],
2615 [
2616 158,
2617 70,
2618 0,
2619 32,
2620 1,
2621 "IMAGE"
2622 ],
2623 [
2624 160,
2625 72,
2626 0,
2627 73,
2628 0,
2629 "INT"
2630 ],
2631 [
2632 161,
2633 72,
2634 1,
2635 73,
2636 1,
2637 "INT"
2638 ],
2639 [
2640 164,
2641 74,
2642 0,
2643 73,
2644 2,
2645 "FLOAT"
2646 ],
2647 [
2648 168,
2649 71,
2650 0,
2651 72,
2652 0,
2653 "IMAGE"
2654 ],
2655 [
2656 169,
2657 73,
2658 1,
2659 31,
2660 3,
2661 "INT"
2662 ],
2663 [
2664 170,
2665 1,
2666 0,
2667 64,
2668 0,
2669 "MODEL"
2670 ],
2671 [
2672 171,
2673 2,
2674 0,
2675 50,
2676 0,
2677 "CLIP"
2678 ]
2679 ],
2680 "groups": [
2681 {
2682 "id": 1,
2683 "title": "SeedVR2, detail improvement/upscale (Optional)",
2684 "bounding": [
2685 2517.987556918573,
2686 300.8396022052916,
2687 1700.892061950401,
2688 720.891004666465
2689 ],
2690 "color": "#3f789e",
2691 "flags": []
2692 },
2693 {
2694 "id": 2,
2695 "title": "Load Models",
2696 "bounding": [
2697 -56.477069239445754,
2698 301.30881096732963,
2699 820.5269848739775,
2700 458.7669928969637
2701 ],
2702 "color": "#A88",
2703 "flags": []
2704 },
2705 {
2706 "id": 4,
2707 "title": "Gen settings",
2708 "bounding": [
2709 765.363324094742,
2710 300.3678327647891,
2711 955.6555488561293,
2712 521.2029016766687
2713 ],
2714 "color": "#3f789e",
2715 "flags": []
2716 },
2717 {
2718 "id": 6,
2719 "title": "Generation & Decode",
2720 "bounding": [
2721 1722.1035748955887,
2722 299.3191815184262,
2723 748.4996121415174,
2724 515.3618843956658
2725 ],
2726 "color": "#3f789e",
2727 "flags": []
2728 },
2729 {
2730 "id": 8,
2731 "title": "Ref Img 2",
2732 "bounding": [
2733 -51.276444655040876,
2734 1149.2926099300792,
2735 403.98016195908804,
2736 359.98557054352796
2737 ],
2738 "color": "#3f789e",
2739 "flags": []
2740 },
2741 {
2742 "id": 9,
2743 "title": "Ref Img 4",
2744 "bounding": [
2745 -50.346603249544785,
2746 1898.2093006739158,
2747 399.27417850158065,
2748 376.30600776809547
2749 ],
2750 "color": "#3f789e",
2751 "flags": []
2752 },
2753 {
2754 "id": 10,
2755 "title": "Ref Img 6",
2756 "bounding": [
2757 355.50713499613937,
2758 763.3068265994915,
2759 389.59024811997847,
2760 382.7374977196504
2761 ],
2762 "color": "#3f789e",
2763 "flags": []
2764 },
2765 {
2766 "id": 11,
2767 "title": "Ref Img 7",
2768 "bounding": [
2769 356.7063179168554,
2770 1152.1282377037212,
2771 396.4815129065377,
2772 358.15204880641636
2773 ],
2774 "color": "#3f789e",
2775 "flags": []
2776 },
2777 {
2778 "id": 12,
2779 "title": "Ref Img 9",
2780 "bounding": [
2781 354.59602350093303,
2782 1900.899229146382,
2783 390.7522708384979,
2784 387.6302249555231
2785 ],
2786 "color": "#3f789e",
2787 "flags": []
2788 },
2789 {
2790 "id": 13,
2791 "title": "Ref Img 8",
2792 "bounding": [
2793 357.85843456607927,
2794 1513.5017388117458,
2795 392.41956836884765,
2796 379.9614348793259
2797 ],
2798 "color": "#3f789e",
2799 "flags": []
2800 },
2801 {
2802 "id": 14,
2803 "title": "Ref Img 5",
2804 "bounding": [
2805 -51.30863574229829,
2806 2275.844055329408,
2807 399.674770544018,
2808 399.09877759641086
2809 ],
2810 "color": "#3f789e",
2811 "flags": []
2812 },
2813 {
2814 "id": 15,
2815 "title": "Ref Img 3",
2816 "bounding": [
2817 -50.42060574898747,
2818 1511.3982360875432,
2819 403.74490801336003,
2820 381.1901727313059
2821 ],
2822 "color": "#3f789e",
2823 "flags": []
2824 },
2825 {
2826 "id": 16,
2827 "title": "Ref Img 1",
2828 "bounding": [
2829 -52.34467073449487,
2830 760.6428792425585,
2831 404.92894800444094,
2832 389.33044766998967
2833 ],
2834 "color": "#3f789e",
2835 "flags": []
2836 },
2837 {
2838 "id": 17,
2839 "title": "Refiner",
2840 "bounding": [
2841 546.9631682502754,
2842 -432.0389298696163,
2843 1197.4217145350226,
2844 728.7799126130972
2845 ],
2846 "color": "#3f789e",
2847 "flags": []
2848 }
2849 ],
2850 "config": [],
2851 "extra": {
2852 "ds": {
2853 "scale": 0.5209868481924399,
2854 "offset": [
2855 1459.4591736415218,
2856 650.3772316983841
2857 ]
2858 },
2859 "frontendVersion": "1.58.5",
2860 "image_studio": {
2861 "release": "v21.0.0",
2862 "source_api": "examples/api/H3_REFERENCE_EDIT_API.json"
2863 },
2864 "VHS_latentpreview": false,
2865 "VHS_latentpreviewrate": 0,
2866 "VHS_MetadataImage": true,
2867 "VHS_KeepIntermediate": true,
2868 "zipTies": [
2869 {
2870 "id": 2,
2871 "pos": [
2872 1207.2808067492067,
2873 446.33818348783876
2874 ],
2875 "angle": -0.042109865618229124,
2876 "links": [
2877 101,
2878 90,
2879 102,
2880 106,
2881 103,
2882 104,
2883 105,
2884 107,
2885 108,
2886 109
2887 ],
2888 "color": "#d6409f"
2889 },
2890 {
2891 "id": 4,
2892 "pos": [
2893 975.9498791625452,
2894 -136.62717799413255
2895 ],
2896 "angle": -1.049885710796137,
2897 "links": [
2898 127,
2899 128,
2900 129,
2901 131,
2902 130,
2903 110,
2904 132,
2905 135,
2906 133,
2907 134
2908 ],
2909 "color": "#d6409f"
2910 },
2911 {
2912 "id": 5,
2913 "pos": [
2914 890.6806625113915,
2915 578.9739125243326
2916 ],
2917 "angle": -1.1173259514489047,
2918 "links": [
2919 127,
2920 128,
2921 129,
2922 130,
2923 131,
2924 110,
2925 90,
2926 101,
2927 132,
2928 135,
2929 102,
2930 133,
2931 134,
2932 103,
2933 106,
2934 104,
2935 105,
2936 107,
2937 109,
2938 108
2939 ],
2940 "color": "#f76b15"
2941 },
2942 {
2943 "id": 6,
2944 "pos": [
2945 802.5527094911258,
2946 1054.3925699056624
2947 ],
2948 "angle": -1.3668634370516246,
2949 "links": [
2950 128,
2951 129,
2952 130,
2953 131,
2954 135,
2955 102,
2956 133,
2957 134,
2958 103,
2959 104,
2960 105,
2961 107,
2962 109,
2963 108
2964 ],
2965 "color": "#d6409f"
2966 },
2967 {
2968 "id": 7,
2969 "pos": [
2970 704.7409742646253,
2971 1580.7898247540973
2972 ],
2973 "angle": -1.258439603000984,
2974 "links": [
2975 129,
2976 130,
2977 131,
2978 135,
2979 103,
2980 104,
2981 105,
2982 109
2983 ],
2984 "color": "#12a594"
2985 },
2986 {
2987 "id": 8,
2988 "pos": [
2989 753.8392081392327,
2990 1285.317477788427
2991 ],
2992 "angle": -1.2944537534263223,
2993 "links": [
2994 128,
2995 129,
2996 130,
2997 131,
2998 135,
2999 102,
3000 134,
3001 103,
3002 104,
3003 105,
3004 109,
3005 108
3006 ],
3007 "color": "#ededed"
3008 }
3009 ]
3010 },
3011 "version": 0.4
3012}