ComfyPilotComfyPilot
工作流 / 图像 / Minimax H3 Image Studio 4k+
Minimax H3 Image Studio 4k+

Minimax H3 Image Studio 4k+

Minimax H3 Image Studio 4k+

MiniMax H3 Image Studio – 4K (Ref2VA + Turbo LoRA + SeedVR2)Text-to-image / reference-based image generation with MiniMax H3, using the video model as a high-quality still generator. Includes an optional SeedVR2 4K refin…

图像 39 个节点 257 次使用 Hopik65 115 KB

下载工作流 JSON 查看官方来源

下载后把文件拖进 ComfyUI 窗口即可打开。

需要的权重

这个工作流引用了 2 个权重文件,其中 0 个能在本站模型库里找到。

minimaxH3ImageStudio_v20Adaptation_3384780.json 模型库暂未收录
minimaxH3ImageStudio_v20Adaptation_3384784.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
ImageBlur × 1
SeedVR2LoadVAEModel × 1
SeedVR2LoadDiTModel × 1
SeedVR2VideoUpscaler × 1
Image Comparer (rgthree) × 1
Power Lora Loader (rgthree) × 1
Fast Groups Muter (rgthree) × 1
MarkdownNote × 1

官方标注的模型

PhotoReal.safetensorsema_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": 69,
5 "last_link_id": 151,
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 84
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 71
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 66
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": 27,
332 "type": "ImageBlur",
333 "pos": [
334 2545.1086046543614,
335 363.12214419184465
336 ],
337 "size": [
338 270,
339 82
340 ],
341 "flags": [],
342 "order": 35,
343 "mode": 2,
344 "inputs": [
345 {
346 "name": "image",
347 "type": "IMAGE",
348 "link": 66
349 }
350 ],
351 "outputs": [
352 {
353 "name": "IMAGE",
354 "type": "IMAGE",
355 "links": [
356 64,
357 65
358 ]
359 }
360 ],
361 "properties": {
362 "cnr_id": "comfy-core",
363 "ver": "0.3.66",
364 "Node name for S&R": "ImageBlur",
365 "ue_properties": {
366 "version": "7.0.1",
367 "widget_ue_connectable": {
368 "blur_radius": true,
369 "sigma": true
370 }
371 }
372 },
373 "widgets_values": [
374 1,
375 1
376 ],
377 "widgets_values_named": {
378 "blur_radius": 1,
379 "sigma": 1
380 }
381 },
382 {
383 "id": 29,
384 "type": "SeedVR2LoadVAEModel",
385 "pos": [
386 2532.6232090380886,
387 575.4828663365603
388 ],
389 "size": [
390 317.9141446387274,
391 317.46834137437736
392 ],
393 "flags": [],
394 "order": 6,
395 "mode": 2,
396 "inputs": [
397 {
398 "name": "torch_compile_args",
399 "shape": 7,
400 "type": "TORCH_COMPILE_ARGS",
401 "link": null
402 }
403 ],
404 "outputs": [
405 {
406 "name": "SEEDVR2_VAE",
407 "type": "SEEDVR2_VAE",
408 "links": [
409 61
410 ]
411 }
412 ],
413 "properties": {
414 "cnr_id": "seedvr2_videoupscaler",
415 "ver": "690cc39379c1481159ddd451368dbf2295930fc6",
416 "Node name for S&R": "SeedVR2LoadVAEModel",
417 "aux_id": "ainvfx/ComfyUI-SeedVR2_VideoUpscaler",
418 "ue_properties": {
419 "version": "7.0.1",
420 "widget_ue_connectable": {
421 "model": true,
422 "device": true,
423 "encode_tiled": true,
424 "encode_tile_size": true,
425 "encode_tile_overlap": true,
426 "decode_tiled": true,
427 "decode_tile_size": true,
428 "decode_tile_overlap": true,
429 "tile_debug": true,
430 "offload_device": true,
431 "cache_model": true
432 }
433 }
434 },
435 "widgets_values": [
436 "ema_vae_fp16.safetensors",
437 "cuda:0",
438 true,
439 256,
440 32,
441 true,
442 256,
443 32,
444 "false",
445 "cpu",
446 false
447 ],
448 "widgets_values_named": {
449 "model": "ema_vae_fp16.safetensors",
450 "device": "cuda:0",
451 "encode_tiled": true,
452 "encode_tile_size": 256,
453 "encode_tile_overlap": 32,
454 "decode_tiled": true,
455 "decode_tile_size": 256,
456 "decode_tile_overlap": 32,
457 "tile_debug": "false",
458 "offload_device": "cpu",
459 "cache_model": false
460 }
461 },
462 {
463 "id": 30,
464 "type": "SeedVR2LoadDiTModel",
465 "pos": [
466 2868.528443513793,
467 360.00174588541984
468 ],
469 "size": [
470 307.6646484375,
471 202
472 ],
473 "flags": [],
474 "order": 7,
475 "mode": 2,
476 "inputs": [
477 {
478 "name": "torch_compile_args",
479 "shape": 7,
480 "type": "TORCH_COMPILE_ARGS",
481 "link": null
482 }
483 ],
484 "outputs": [
485 {
486 "name": "SEEDVR2_DIT",
487 "type": "SEEDVR2_DIT",
488 "links": [
489 60
490 ]
491 }
492 ],
493 "properties": {
494 "cnr_id": "seedvr2_videoupscaler",
495 "ver": "690cc39379c1481159ddd451368dbf2295930fc6",
496 "Node name for S&R": "SeedVR2LoadDiTModel",
497 "aux_id": "ainvfx/ComfyUI-SeedVR2_VideoUpscaler",
498 "ue_properties": {
499 "version": "7.0.1",
500 "widget_ue_connectable": {
501 "model": true,
502 "device": true,
503 "blocks_to_swap": true,
504 "swap_io_components": true,
505 "offload_device": true,
506 "cache_model": true,
507 "attention_mode": true
508 }
509 }
510 },
511 "widgets_values": [
512 "seedvr2_ema_7b-Q4_K_M.gguf",
513 "cuda:0",
514 36,
515 true,
516 "cpu",
517 false,
518 "sdpa"
519 ],
520 "widgets_values_named": {
521 "model": "seedvr2_ema_7b-Q4_K_M.gguf",
522 "device": "cuda:0",
523 "blocks_to_swap": 36,
524 "swap_io_components": true,
525 "offload_device": "cpu",
526 "cache_model": false,
527 "attention_mode": "sdpa"
528 }
529 },
530 {
531 "id": 31,
532 "type": "SeedVR2VideoUpscaler",
533 "pos": [
534 2875.3684006393396,
535 608.7454495286365
536 ],
537 "size": [
538 289.06908246036255,
539 386
540 ],
541 "flags": [],
542 "order": 36,
543 "mode": 2,
544 "inputs": [
545 {
546 "name": "image",
547 "type": "IMAGE",
548 "link": 64
549 },
550 {
551 "name": "dit",
552 "type": "SEEDVR2_DIT",
553 "link": 60
554 },
555 {
556 "name": "vae",
557 "type": "SEEDVR2_VAE",
558 "link": 61
559 }
560 ],
561 "outputs": [
562 {
563 "name": "IMAGE",
564 "type": "IMAGE",
565 "links": [
566 63,
567 83
568 ]
569 }
570 ],
571 "properties": {
572 "cnr_id": "seedvr2_videoupscaler",
573 "ver": "690cc39379c1481159ddd451368dbf2295930fc6",
574 "Node name for S&R": "SeedVR2VideoUpscaler",
575 "aux_id": "ainvfx/ComfyUI-SeedVR2_VideoUpscaler",
576 "ue_properties": {
577 "version": "7.0.1",
578 "widget_ue_connectable": {
579 "seed": true,
580 "resolution": true,
581 "max_resolution": true,
582 "batch_size": true,
583 "uniform_batch_size": true,
584 "color_correction": true,
585 "temporal_overlap": true,
586 "prepend_frames": true,
587 "input_noise_scale": true,
588 "latent_noise_scale": true,
589 "offload_device": true,
590 "enable_debug": true
591 }
592 }
593 },
594 "widgets_values": [
595 1102,
596 "fixed",
597 4096,
598 4096,
599 1,
600 false,
601 "lab",
602 0,
603 0,
604 0,
605 0,
606 "cpu",
607 false
608 ],
609 "widgets_values_named": {
610 "seed": 1102,
611 "control_after_generate": "fixed",
612 "resolution": 4096,
613 "max_resolution": 4096,
614 "batch_size": 1,
615 "uniform_batch_size": false,
616 "color_correction": "lab",
617 "temporal_overlap": 0,
618 "prepend_frames": 0,
619 "input_noise_scale": 0,
620 "latent_noise_scale": 0,
621 "offload_device": "cpu",
622 "enable_debug": false
623 }
624 },
625 {
626 "id": 32,
627 "type": "Image Comparer (rgthree)",
628 "pos": [
629 3195.942387983182,
630 374.95318352421475
631 ],
632 "size": [
633 513.1261135479682,
634 478.9293805454947
635 ],
636 "flags": [],
637 "order": 37,
638 "mode": 0,
639 "inputs": [
640 {
641 "dir": 3,
642 "name": "image_a",
643 "type": "IMAGE",
644 "link": 65
645 },
646 {
647 "dir": 3,
648 "name": "image_b",
649 "type": "IMAGE",
650 "link": 63
651 }
652 ],
653 "outputs": [
654 {
655 "dir": 4,
656 "name": "images",
657 "shape": 3,
658 "type": "IMAGE",
659 "links": []
660 }
661 ],
662 "title": "Image Comparer (rgthree) 7BQ8",
663 "properties": {
664 "cnr_id": "rgthree-comfy",
665 "ver": "1.0.2508012353",
666 "comparer_mode": "Slide",
667 "ue_properties": {
668 "version": "7.0.1",
669 "widget_ue_connectable": []
670 }
671 },
672 "widgets_values": [
673 [
674 {
675 "name": "A",
676 "selected": true,
677 "url": "/api/view?filename=rgthree.compare._temp_nxarc_00005_.png&type=temp&subfolder=&rand=0.9991026376243927"
678 },
679 {
680 "name": "B",
681 "selected": true,
682 "url": "/api/view?filename=rgthree.compare._temp_nxarc_00006_.png&type=temp&subfolder=&rand=0.6970344726588816"
683 }
684 ]
685 ],
686 "widgets_values_named": {
687 "rgthree_comparer": {
688 "images": [
689 {
690 "name": "A",
691 "selected": true,
692 "url": "/api/view?filename=rgthree.compare._temp_nxarc_00005_.png&type=temp&subfolder=&rand=0.9991026376243927",
693 "img": []
694 },
695 {
696 "name": "B",
697 "selected": true,
698 "url": "/api/view?filename=rgthree.compare._temp_nxarc_00006_.png&type=temp&subfolder=&rand=0.6970344726588816",
699 "img": []
700 }
701 ]
702 }
703 }
704 },
705 {
706 "id": 33,
707 "type": "Power Lora Loader (rgthree)",
708 "pos": [
709 381.9922824102935,
710 363.82016226506164
711 ],
712 "size": [
713 344.0720418482439,
714 153.91440241025703
715 ],
716 "flags": [],
717 "order": 25,
718 "mode": 0,
719 "inputs": [
720 {
721 "dir": 3,
722 "name": "model",
723 "type": "MODEL",
724 "link": 84
725 },
726 {
727 "dir": 3,
728 "name": "clip",
729 "type": "CLIP",
730 "link": 71
731 }
732 ],
733 "outputs": [
734 {
735 "dir": 4,
736 "name": "MODEL",
737 "shape": 3,
738 "type": "MODEL",
739 "links": [
740 140
741 ]
742 },
743 {
744 "dir": 4,
745 "name": "CLIP",
746 "shape": 3,
747 "type": "CLIP",
748 "links": [
749 95
750 ]
751 }
752 ],
753 "properties": {
754 "cnr_id": "rgthree-comfy",
755 "ver": "1.0.2610300006",
756 "Show Strengths": "Single Strength",
757 "Match": ""
758 },
759 "widgets_values": [
760 [],
761 {
762 "type": "PowerLoraLoaderHeaderWidget"
763 },
764 {
765 "on": false,
766 "lora": "PhotoReal.safetensors",
767 "strength": 0.8,
768 "strengthTwo": null
769 },
770 [],
771 ""
772 ],
773 "widgets_values_named": {
774 "divider": [],
775 "PowerLoraLoaderHeaderWidget": {
776 "type": "PowerLoraLoaderHeaderWidget"
777 },
778 "lora_1": {
779 "on": false,
780 "lora": "PhotoReal.safetensors",
781 "strength": 0.8,
782 "strengthTwo": null
783 },
784 "divider#2": [],
785 "➕ Add Lora": ""
786 },
787 "color": "#322",
788 "bgcolor": "#533"
789 },
790 {
791 "id": 34,
792 "type": "LoadImage",
793 "pos": [
794 -50.94572574862306,
795 2341.2521717648697
796 ],
797 "size": [
798 390,
799 314
800 ],
801 "flags": {
802 "collapsed": false
803 },
804 "order": 8,
805 "mode": 2,
806 "inputs": [],
807 "outputs": [
808 {
809 "name": "IMAGE",
810 "type": "IMAGE",
811 "links": [
812 105,
813 131
814 ]
815 },
816 {
817 "name": "MASK",
818 "type": "MASK",
819 "links": []
820 }
821 ],
822 "title": "Image 5",
823 "properties": {
824 "cnr_id": "comfy-core",
825 "ver": "0.34.0",
826 "Node name for S&R": "LoadImage"
827 },
828 "widgets_values": [
829 "images.jpg",
830 "image"
831 ],
832 "widgets_values_named": {
833 "image": "images.jpg",
834 "upload": "image"
835 }
836 },
837 {
838 "id": 35,
839 "type": "LoadImage",
840 "pos": [
841 -47.19353927811031,
842 1567.276163971839
843 ],
844 "size": [
845 390,
846 314
847 ],
848 "flags": {
849 "collapsed": false
850 },
851 "order": 9,
852 "mode": 2,
853 "inputs": [],
854 "outputs": [
855 {
856 "name": "IMAGE",
857 "type": "IMAGE",
858 "links": [
859 103,
860 129
861 ]
862 },
863 {
864 "name": "MASK",
865 "type": "MASK",
866 "links": []
867 }
868 ],
869 "title": "Image 3",
870 "properties": {
871 "cnr_id": "comfy-core",
872 "ver": "0.34.0",
873 "Node name for S&R": "LoadImage"
874 },
875 "widgets_values": [
876 "images.jpg",
877 "image"
878 ],
879 "widgets_values_named": {
880 "image": "images.jpg",
881 "upload": "image"
882 }
883 },
884 {
885 "id": 36,
886 "type": "LoadImage",
887 "pos": [
888 -45.20338044027696,
889 1960.4846883931234
890 ],
891 "size": [
892 387.5536363820635,
893 314
894 ],
895 "flags": {
896 "collapsed": false
897 },
898 "order": 10,
899 "mode": 2,
900 "inputs": [],
901 "outputs": [
902 {
903 "name": "IMAGE",
904 "type": "IMAGE",
905 "links": [
906 104,
907 130
908 ]
909 },
910 {
911 "name": "MASK",
912 "type": "MASK",
913 "links": []
914 }
915 ],
916 "title": "Image 4",
917 "properties": {
918 "cnr_id": "comfy-core",
919 "ver": "0.34.0",
920 "Node name for S&R": "LoadImage"
921 },
922 "widgets_values": [
923 "images.jpg",
924 "image"
925 ],
926 "widgets_values_named": {
927 "image": "images.jpg",
928 "upload": "image"
929 }
930 },
931 {
932 "id": 37,
933 "type": "LoadImage",
934 "pos": [
935 350.87608738956527,
936 828.0811610534995
937 ],
938 "size": [
939 390,
940 314
941 ],
942 "flags": {
943 "collapsed": false
944 },
945 "order": 11,
946 "mode": 2,
947 "inputs": [],
948 "outputs": [
949 {
950 "name": "IMAGE",
951 "type": "IMAGE",
952 "links": [
953 106,
954 132
955 ]
956 },
957 {
958 "name": "MASK",
959 "type": "MASK",
960 "links": []
961 }
962 ],
963 "title": "Image 6",
964 "properties": {
965 "cnr_id": "comfy-core",
966 "ver": "0.34.0",
967 "Node name for S&R": "LoadImage"
968 },
969 "widgets_values": [
970 "images.jpg",
971 "image"
972 ],
973 "widgets_values_named": {
974 "image": "images.jpg",
975 "upload": "image"
976 }
977 },
978 {
979 "id": 38,
980 "type": "LoadImage",
981 "pos": [
982 359.7685359160873,
983 1205.0749340183063
984 ],
985 "size": [
986 392.01824998479765,
987 311.3090000202699
988 ],
989 "flags": {
990 "collapsed": false
991 },
992 "order": 12,
993 "mode": 2,
994 "inputs": [],
995 "outputs": [
996 {
997 "name": "IMAGE",
998 "type": "IMAGE",
999 "links": [
1000 107,
1001 133
1002 ]
1003 },
1004 {
1005 "name": "MASK",
1006 "type": "MASK",
1007 "links": []
1008 }
1009 ],
1010 "title": "Image 7",
1011 "properties": {
1012 "cnr_id": "comfy-core",
1013 "ver": "0.34.0",
1014 "Node name for S&R": "LoadImage"
1015 },
1016 "widgets_values": [
1017 "images.jpg",
1018 "image"
1019 ],
1020 "widgets_values_named": {
1021 "image": "images.jpg",
1022 "upload": "image"
1023 }
1024 },
1025 {
1026 "id": 39,
1027 "type": "LoadImage",
1028 "pos": [
1029 361.9014769139223,
1030 1569.7421062596138
1031 ],
1032 "size": [
1033 390,
1034 314
1035 ],
1036 "flags": {
1037 "collapsed": false
1038 },
1039 "order": 13,
1040 "mode": 2,
1041 "inputs": [],
1042 "outputs": [
1043 {
1044 "name": "IMAGE",
1045 "type": "IMAGE",
1046 "links": [
1047 108,
1048 134
1049 ]
1050 },
1051 {
1052 "name": "MASK",
1053 "type": "MASK",
1054 "links": []
1055 }
1056 ],
1057 "title": "Image 8",
1058 "properties": {
1059 "cnr_id": "comfy-core",
1060 "ver": "0.34.0",
1061 "Node name for S&R": "LoadImage"
1062 },
1063 "widgets_values": [
1064 "images.jpg",
1065 "image"
1066 ],
1067 "widgets_values_named": {
1068 "image": "images.jpg",
1069 "upload": "image"
1070 }
1071 },
1072 {
1073 "id": 40,
1074 "type": "LoadImage",
1075 "pos": [
1076 354.6299140588164,
1077 1954.0273752515875
1078 ],
1079 "size": [
1080 390,
1081 314
1082 ],
1083 "flags": {
1084 "collapsed": false
1085 },
1086 "order": 14,
1087 "mode": 2,
1088 "inputs": [],
1089 "outputs": [
1090 {
1091 "name": "IMAGE",
1092 "type": "IMAGE",
1093 "links": [
1094 109,
1095 135
1096 ]
1097 },
1098 {
1099 "name": "MASK",
1100 "type": "MASK",
1101 "links": []
1102 }
1103 ],
1104 "title": "Image 9",
1105 "properties": {
1106 "cnr_id": "comfy-core",
1107 "ver": "0.34.0",
1108 "Node name for S&R": "LoadImage"
1109 },
1110 "widgets_values": [
1111 "images.jpg",
1112 "image"
1113 ],
1114 "widgets_values_named": {
1115 "image": "images.jpg",
1116 "upload": "image"
1117 }
1118 },
1119 {
1120 "id": 41,
1121 "type": "Fast Groups Muter (rgthree)",
1122 "pos": [
1123 1279.242965820201,
1124 890.3021981018636
1125 ],
1126 "size": [
1127 293.84809166352613,
1128 394
1129 ],
1130 "flags": [],
1131 "order": 15,
1132 "mode": 0,
1133 "inputs": [],
1134 "outputs": [
1135 {
1136 "name": "OPT_CONNECTION",
1137 "type": "*",
1138 "links": null
1139 }
1140 ],
1141 "properties": {
1142 "matchColors": "",
1143 "matchTitle": "",
1144 "showNav": true,
1145 "showAllGraphs": true,
1146 "sort": "position",
1147 "customSortAlphabet": "",
1148 "toggleRestriction": "default"
1149 }
1150 },
1151 {
1152 "id": 43,
1153 "type": "SaveImage",
1154 "pos": [
1155 3742.1250327109683,
1156 393.7452995059146
1157 ],
1158 "size": [
1159 390,
1160 270
1161 ],
1162 "flags": [],
1163 "order": 38,
1164 "mode": 2,
1165 "inputs": [
1166 {
1167 "name": "images",
1168 "type": "IMAGE",
1169 "link": 83
1170 }
1171 ],
1172 "outputs": [
1173 {
1174 "name": "images",
1175 "type": "IMAGE",
1176 "links": []
1177 }
1178 ],
1179 "properties": {
1180 "cnr_id": "comfy-core",
1181 "ver": "0.34.0",
1182 "Node name for S&R": "SaveImage"
1183 },
1184 "widgets_values": [
1185 "MiniMax_H3_Image"
1186 ],
1187 "widgets_values_named": {
1188 "filename_prefix": "MiniMax_H3_Image"
1189 }
1190 },
1191 {
1192 "id": 44,
1193 "type": "PrimitiveStringMultiline",
1194 "pos": [
1195 739.7400444609955,
1196 880.388412000533
1197 ],
1198 "size": [
1199 480,
1200 660
1201 ],
1202 "flags": {
1203 "collapsed": false
1204 },
1205 "order": 16,
1206 "mode": 0,
1207 "inputs": [],
1208 "outputs": [
1209 {
1210 "name": "STRING",
1211 "type": "STRING",
1212 "links": [
1213 143
1214 ]
1215 }
1216 ],
1217 "title": "Text (Prompt)",
1218 "properties": {
1219 "cnr_id": "comfy-core",
1220 "ver": "0.44.0",
1221 "Node name for S&R": "PrimitiveStringMultiline"
1222 },
1223 "widgets_values": [
1224 ""
1225 ],
1226 "widgets_values_named": {
1227 "value": ""
1228 },
1229 "color": "#223",
1230 "bgcolor": "#335"
1231 },
1232 {
1233 "id": 45,
1234 "type": "MarkdownNote",
1235 "pos": [
1236 -603.4814273539805,
1237 341.92276428631504
1238 ],
1239 "size": [
1240 520,
1241 770
1242 ],
1243 "flags": [],
1244 "order": 17,
1245 "mode": 0,
1246 "inputs": [],
1247 "outputs": [],
1248 "properties": [],
1249 "widgets_values": [
1250 "## 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**: 79.63 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"
1251 ],
1252 "widgets_values_named": {
1253 "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**: 79.63 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"
1254 },
1255 "color": "#222",
1256 "bgcolor": "#000"
1257 },
1258 {
1259 "id": 47,
1260 "type": "KSampler",
1261 "pos": [
1262 1794.6595299688718,
1263 394.13734950187495
1264 ],
1265 "size": [
1266 301.962890625,
1267 474
1268 ],
1269 "flags": [],
1270 "order": 32,
1271 "mode": 0,
1272 "inputs": [
1273 {
1274 "name": "model",
1275 "type": "MODEL",
1276 "link": 141
1277 },
1278 {
1279 "name": "positive",
1280 "type": "CONDITIONING",
1281 "link": 99
1282 },
1283 {
1284 "name": "negative",
1285 "type": "CONDITIONING",
1286 "link": 98
1287 },
1288 {
1289 "name": "latent_image",
1290 "type": "LATENT",
1291 "link": 138
1292 }
1293 ],
1294 "outputs": [
1295 {
1296 "name": "LATENT",
1297 "type": "LATENT",
1298 "links": [
1299 89
1300 ]
1301 }
1302 ],
1303 "properties": {
1304 "cnr_id": "comfy-core",
1305 "ver": "0.45.0",
1306 "Node name for S&R": "KSampler"
1307 },
1308 "widgets_values": [
1309 269837239392791,
1310 "randomize",
1311 6,
1312 1,
1313 "euler",
1314 "simple",
1315 1
1316 ],
1317 "widgets_values_named": {
1318 "seed": 269837239392791,
1319 "control_after_generate": "randomize",
1320 "steps": 6,
1321 "cfg": 1,
1322 "sampler_name": "euler",
1323 "scheduler": "simple",
1324 "denoise": 1
1325 }
1326 },
1327 {
1328 "id": 49,
1329 "type": "VAEDecode",
1330 "pos": [
1331 2183.130024456669,
1332 349.5784021185119
1333 ],
1334 "size": [
1335 174.439453125,
1336 46
1337 ],
1338 "flags": [],
1339 "order": 33,
1340 "mode": 0,
1341 "inputs": [
1342 {
1343 "name": "samples",
1344 "type": "LATENT",
1345 "link": 89
1346 },
1347 {
1348 "name": "vae",
1349 "type": "VAE",
1350 "link": 90
1351 }
1352 ],
1353 "outputs": [
1354 {
1355 "name": "IMAGE",
1356 "type": "IMAGE",
1357 "links": [
1358 100
1359 ]
1360 }
1361 ],
1362 "properties": {
1363 "cnr_id": "comfy-core",
1364 "ver": "0.45.0",
1365 "Node name for S&R": "VAEDecode"
1366 }
1367 },
1368 {
1369 "id": 50,
1370 "type": "MiniMaxH3ImageConditioning",
1371 "pos": [
1372 1268.8557577401202,
1373 372.4492661835387
1374 ],
1375 "size": [
1376 440,
1377 436
1378 ],
1379 "flags": [],
1380 "order": 30,
1381 "mode": 0,
1382 "inputs": [
1383 {
1384 "name": "clip",
1385 "type": "CLIP",
1386 "link": 95
1387 },
1388 {
1389 "name": "vae",
1390 "shape": 7,
1391 "type": "VAE",
1392 "link": null
1393 },
1394 {
1395 "name": "images.image_0",
1396 "shape": 7,
1397 "type": "IMAGE",
1398 "link": 101
1399 },
1400 {
1401 "name": "images.image_1",
1402 "shape": 7,
1403 "type": "IMAGE",
1404 "link": 102
1405 },
1406 {
1407 "name": "prompt",
1408 "type": "STRING",
1409 "widget": {
1410 "name": "prompt"
1411 },
1412 "link": 151
1413 },
1414 {
1415 "name": "width",
1416 "type": "INT",
1417 "widget": {
1418 "name": "width"
1419 },
1420 "link": 92
1421 },
1422 {
1423 "name": "height",
1424 "type": "INT",
1425 "widget": {
1426 "name": "height"
1427 },
1428 "link": 93
1429 },
1430 {
1431 "name": "images.image_10",
1432 "shape": 7,
1433 "type": "IMAGE",
1434 "link": null
1435 },
1436 {
1437 "name": "images.image_11",
1438 "shape": 7,
1439 "type": "IMAGE",
1440 "link": null
1441 },
1442 {
1443 "name": "images.image_12",
1444 "shape": 7,
1445 "type": "IMAGE",
1446 "link": null
1447 },
1448 {
1449 "name": "images.image_2",
1450 "shape": 7,
1451 "type": "IMAGE",
1452 "link": 103
1453 },
1454 {
1455 "name": "images.image_3",
1456 "shape": 7,
1457 "type": "IMAGE",
1458 "link": 104
1459 },
1460 {
1461 "name": "images.image_4",
1462 "shape": 7,
1463 "type": "IMAGE",
1464 "link": 105
1465 },
1466 {
1467 "name": "images.image_5",
1468 "shape": 7,
1469 "type": "IMAGE",
1470 "link": 106
1471 },
1472 {
1473 "name": "images.image_6",
1474 "shape": 7,
1475 "type": "IMAGE",
1476 "link": 107
1477 },
1478 {
1479 "name": "images.image_7",
1480 "shape": 7,
1481 "type": "IMAGE",
1482 "link": 108
1483 },
1484 {
1485 "name": "images.image_8",
1486 "shape": 7,
1487 "type": "IMAGE",
1488 "link": 109
1489 },
1490 {
1491 "name": "images.image_9",
1492 "shape": 7,
1493 "type": "IMAGE",
1494 "link": null
1495 }
1496 ],
1497 "outputs": [
1498 {
1499 "name": "positive",
1500 "type": "CONDITIONING",
1501 "links": [
1502 96,
1503 99
1504 ]
1505 },
1506 {
1507 "name": "LATENT",
1508 "type": "LATENT",
1509 "links": [
1510 138
1511 ]
1512 }
1513 ],
1514 "properties": {
1515 "cnr_id": "comfy-core",
1516 "ver": "0.45.0",
1517 "Node name for S&R": "MiniMaxH3ImageConditioning"
1518 },
1519 "widgets_values": [
1520 "",
1521 3104,
1522 3104
1523 ],
1524 "widgets_values_named": {
1525 "prompt": "",
1526 "width": 3104,
1527 "height": 3104
1528 }
1529 },
1530 {
1531 "id": 51,
1532 "type": "ConditioningZeroOut",
1533 "pos": [
1534 1811.6309184950376,
1535 442.3786301675202
1536 ],
1537 "size": [
1538 229.5767578125,
1539 26
1540 ],
1541 "flags": [],
1542 "order": 31,
1543 "mode": 0,
1544 "inputs": [
1545 {
1546 "name": "conditioning",
1547 "type": "CONDITIONING",
1548 "link": 96
1549 }
1550 ],
1551 "outputs": [
1552 {
1553 "name": "CONDITIONING",
1554 "type": "CONDITIONING",
1555 "links": [
1556 98
1557 ]
1558 }
1559 ],
1560 "properties": {
1561 "cnr_id": "comfy-core",
1562 "ver": "0.45.0",
1563 "Node name for S&R": "ConditioningZeroOut"
1564 }
1565 },
1566 {
1567 "id": 52,
1568 "type": "Note",
1569 "pos": [
1570 1611.157473174652,
1571 914.1818785324105
1572 ],
1573 "size": [
1574 295.10287435896885,
1575 139.10867190411932
1576 ],
1577 "flags": [],
1578 "order": 18,
1579 "mode": 0,
1580 "inputs": [],
1581 "outputs": [],
1582 "properties": [],
1583 "widgets_values": [
1584 "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."
1585 ],
1586 "widgets_values_named": {
1587 "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."
1588 },
1589 "color": "#432",
1590 "bgcolor": "#653"
1591 },
1592 {
1593 "id": 54,
1594 "type": "PreviewAny",
1595 "pos": [
1596 1386.8408378099425,
1597 -204.51726590007362
1598 ],
1599 "size": [
1600 328.90005808500405,
1601 427.2830853849557
1602 ],
1603 "flags": [],
1604 "order": 29,
1605 "mode": 4,
1606 "inputs": [
1607 {
1608 "name": "source",
1609 "type": "*",
1610 "link": 144
1611 }
1612 ],
1613 "outputs": [
1614 {
1615 "name": "STRING",
1616 "type": "STRING",
1617 "links": [
1618 151
1619 ]
1620 }
1621 ],
1622 "properties": {
1623 "cnr_id": "comfy-core",
1624 "ver": "0.45.0",
1625 "Node name for S&R": "PreviewAny"
1626 },
1627 "widgets_values": [],
1628 "widgets_values_named": []
1629 },
1630 {
1631 "id": 55,
1632 "type": "TextGenerateV2",
1633 "pos": [
1634 1095.8958200600923,
1635 -358.901275920359
1636 ],
1637 "size": [
1638 283.0760508807234,
1639 716
1640 ],
1641 "flags": [],
1642 "order": 26,
1643 "mode": 4,
1644 "showAdvanced": true,
1645 "inputs": [
1646 {
1647 "name": "clip",
1648 "type": "CLIP",
1649 "link": 110
1650 },
1651 {
1652 "name": "images.image1",
1653 "shape": 7,
1654 "type": "IMAGE",
1655 "link": 127
1656 },
1657 {
1658 "name": "videos.video1",
1659 "shape": 7,
1660 "type": "IMAGE",
1661 "link": null
1662 },
1663 {
1664 "name": "audios.audio1",
1665 "shape": 7,
1666 "type": "AUDIO",
1667 "link": null
1668 },
1669 {
1670 "name": "system_prompt",
1671 "shape": 7,
1672 "type": "STRING",
1673 "link": 112
1674 },
1675 {
1676 "name": "prompt",
1677 "type": "STRING",
1678 "widget": {
1679 "name": "prompt"
1680 },
1681 "link": 149
1682 },
1683 {
1684 "name": "images.image3",
1685 "shape": 7,
1686 "type": "IMAGE",
1687 "link": 129
1688 },
1689 {
1690 "name": "images.image4",
1691 "shape": 7,
1692 "type": "IMAGE",
1693 "link": 130
1694 },
1695 {
1696 "name": "images.image5",
1697 "shape": 7,
1698 "type": "IMAGE",
1699 "link": 131
1700 },
1701 {
1702 "name": "images.image6",
1703 "shape": 7,
1704 "type": "IMAGE",
1705 "link": 132
1706 },
1707 {
1708 "name": "images.image7",
1709 "shape": 7,
1710 "type": "IMAGE",
1711 "link": 133
1712 },
1713 {
1714 "name": "images.image8",
1715 "shape": 7,
1716 "type": "IMAGE",
1717 "link": 134
1718 },
1719 {
1720 "name": "images.image9",
1721 "shape": 7,
1722 "type": "IMAGE",
1723 "link": 135
1724 },
1725 {
1726 "name": "images.image10",
1727 "shape": 7,
1728 "type": "IMAGE",
1729 "link": null
1730 },
1731 {
1732 "name": "images.image11",
1733 "shape": 7,
1734 "type": "IMAGE",
1735 "link": null
1736 },
1737 {
1738 "name": "images.image2",
1739 "shape": 7,
1740 "type": "IMAGE",
1741 "link": 128
1742 },
1743 {
1744 "name": "audios.audio2",
1745 "shape": 7,
1746 "type": "AUDIO",
1747 "link": null
1748 },
1749 {
1750 "name": "audios.audio3",
1751 "shape": 7,
1752 "type": "AUDIO",
1753 "link": null
1754 }
1755 ],
1756 "outputs": [
1757 {
1758 "name": "generated_text",
1759 "type": "STRING",
1760 "links": [
1761 142
1762 ]
1763 },
1764 {
1765 "name": "thinking",
1766 "type": "STRING",
1767 "links": []
1768 }
1769 ],
1770 "properties": {
1771 "cnr_id": "comfy-core",
1772 "ver": "0.45.0",
1773 "Node name for S&R": "TextGenerateV2"
1774 },
1775 "widgets_values": [
1776 "",
1777 8196,
1778 "on",
1779 1,
1780 20,
1781 0.95,
1782 0,
1783 1,
1784 0,
1785 0,
1786 true,
1787 true,
1788 "auto"
1789 ],
1790 "widgets_values_named": {
1791 "prompt": "",
1792 "max_length": 8196,
1793 "sampling_mode": "on",
1794 "sampling_mode.temperature": 1,
1795 "sampling_mode.top_k": 20,
1796 "sampling_mode.top_p": 0.95,
1797 "sampling_mode.min_p": 0,
1798 "sampling_mode.repetition_penalty": 1,
1799 "sampling_mode.presence_penalty": 0,
1800 "sampling_mode.seed": 0,
1801 "thinking": true,
1802 "use_default_template": true,
1803 "mtp": "auto"
1804 }
1805 },
1806 {
1807 "id": 56,
1808 "type": "CLIPLoader",
1809 "pos": [
1810 567.6513796256692,
1811 185.0141427322102
1812 ],
1813 "size": [
1814 347.77795614400543,
1815 106
1816 ],
1817 "flags": [],
1818 "order": 19,
1819 "mode": 4,
1820 "showAdvanced": false,
1821 "inputs": [],
1822 "outputs": [
1823 {
1824 "name": "CLIP",
1825 "type": "CLIP",
1826 "links": [
1827 110
1828 ]
1829 }
1830 ],
1831 "title": "Загрузить CLIP (Refiner)",
1832 "properties": {
1833 "cnr_id": "comfy-core",
1834 "ver": "0.45.0",
1835 "Node name for S&R": "CLIPLoader",
1836 "models": [
1837 {
1838 "name": "qwen3.5_9b_w4a8_gptq.safetensors",
1839 "url": "https://huggingface.co/Comfy-Org/Qwen3.5/resolve/main/text_encoders/qwen3.5_9b_w4a8_gptq.safetensors",
1840 "directory": "text_encoders"
1841 }
1842 ]
1843 },
1844 "widgets_values": [
1845 "qwen3.5_9b_w4a8_gptq.safetensors",
1846 "minimax",
1847 "default"
1848 ],
1849 "widgets_values_named": {
1850 "clip_name": "qwen3.5_9b_w4a8_gptq.safetensors",
1851 "type": "minimax",
1852 "device": "default"
1853 },
1854 "color": "#322",
1855 "bgcolor": "#533"
1856 },
1857 {
1858 "id": 57,
1859 "type": "PrimitiveStringMultiline",
1860 "pos": [
1861 548.4152565205796,
1862 -365.83400192098753
1863 ],
1864 "size": [
1865 517.796416818221,
1866 143.22663572666892
1867 ],
1868 "flags": [],
1869 "order": 20,
1870 "mode": 4,
1871 "inputs": [],
1872 "outputs": [
1873 {
1874 "name": "STRING",
1875 "type": "STRING",
1876 "links": [
1877 112
1878 ]
1879 }
1880 ],
1881 "title": "Text (system_prompt)",
1882 "properties": {
1883 "cnr_id": "comfy-core",
1884 "ver": "0.45.0",
1885 "Node name for S&R": "PrimitiveStringMultiline"
1886 },
1887 "widgets_values": [
1888 "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."
1889 ],
1890 "widgets_values_named": {
1891 "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."
1892 }
1893 },
1894 {
1895 "id": 64,
1896 "type": "MiniMaxH3SigmaShift",
1897 "pos": [
1898 891.7152018659292,
1899 583.3831364454603
1900 ],
1901 "size": [
1902 270,
1903 82
1904 ],
1905 "flags": [],
1906 "order": 27,
1907 "mode": 0,
1908 "inputs": [
1909 {
1910 "name": "model",
1911 "type": "MODEL",
1912 "link": 140
1913 }
1914 ],
1915 "outputs": [
1916 {
1917 "name": "MODEL",
1918 "type": "MODEL",
1919 "links": [
1920 141
1921 ]
1922 }
1923 ],
1924 "properties": {
1925 "cnr_id": "comfy-core",
1926 "ver": "0.45.0",
1927 "Node name for S&R": "MiniMaxH3SigmaShift"
1928 },
1929 "widgets_values": [
1930 3,
1931 100
1932 ],
1933 "widgets_values_named": {
1934 "shift_video": 3,
1935 "shift_audio": 100
1936 }
1937 },
1938 {
1939 "id": 65,
1940 "type": "Note",
1941 "pos": [
1942 1933.4929360238134,
1943 914.124505271675
1944 ],
1945 "size": [
1946 295.10287435896885,
1947 139.10867190411932
1948 ],
1949 "flags": [],
1950 "order": 21,
1951 "mode": 0,
1952 "inputs": [],
1953 "outputs": [],
1954 "properties": [],
1955 "widgets_values": [
1956 "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."
1957 ],
1958 "widgets_values_named": {
1959 "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."
1960 },
1961 "color": "#432",
1962 "bgcolor": "#653"
1963 },
1964 {
1965 "id": 66,
1966 "type": "PrimitiveBoolean",
1967 "pos": [
1968 889.9198423696647,
1969 727.6527619461102
1970 ],
1971 "size": [
1972 270,
1973 58
1974 ],
1975 "flags": [],
1976 "order": 22,
1977 "mode": 0,
1978 "inputs": [],
1979 "outputs": [
1980 {
1981 "name": "BOOLEAN",
1982 "type": "BOOLEAN",
1983 "links": [
1984 147
1985 ]
1986 }
1987 ],
1988 "title": "Enable Promp Refiner",
1989 "properties": {
1990 "cnr_id": "comfy-core",
1991 "ver": "0.45.0",
1992 "Node name for S&R": "PrimitiveBoolean"
1993 },
1994 "widgets_values": [
1995 false
1996 ],
1997 "widgets_values_named": {
1998 "value": false
1999 }
2000 },
2001 {
2002 "id": 67,
2003 "type": "ComfySwitchNode",
2004 "pos": [
2005 1402.8939201257251,
2006 -325.4806557512387
2007 ],
2008 "size": [
2009 270,
2010 78
2011 ],
2012 "flags": [],
2013 "order": 28,
2014 "mode": 4,
2015 "inputs": [
2016 {
2017 "name": "on_false",
2018 "shape": 7,
2019 "type": "STRING",
2020 "link": 143
2021 },
2022 {
2023 "name": "on_true",
2024 "shape": 7,
2025 "type": "STRING",
2026 "link": 142
2027 },
2028 {
2029 "name": "switch",
2030 "type": "BOOLEAN",
2031 "widget": {
2032 "name": "switch"
2033 },
2034 "link": 147
2035 }
2036 ],
2037 "outputs": [
2038 {
2039 "name": "output",
2040 "type": "STRING",
2041 "links": [
2042 144
2043 ]
2044 }
2045 ],
2046 "properties": {
2047 "cnr_id": "comfy-core",
2048 "ver": "0.45.0",
2049 "Node name for S&R": "ComfySwitchNode"
2050 },
2051 "widgets_values": [
2052 false
2053 ],
2054 "widgets_values_named": {
2055 "switch": false
2056 }
2057 },
2058 {
2059 "id": 68,
2060 "type": "PrimitiveStringMultiline",
2061 "pos": [
2062 548.7645818740514,
2063 -179.61207713992954
2064 ],
2065 "size": [
2066 489.41849992905577,
2067 189.0750035472086
2068 ],
2069 "flags": {
2070 "collapsed": false
2071 },
2072 "order": 23,
2073 "mode": 4,
2074 "inputs": [],
2075 "outputs": [
2076 {
2077 "name": "STRING",
2078 "type": "STRING",
2079 "links": [
2080 149
2081 ]
2082 }
2083 ],
2084 "title": "Text (Prompt)",
2085 "properties": {
2086 "cnr_id": "comfy-core",
2087 "ver": "0.44.0",
2088 "Node name for S&R": "PrimitiveStringMultiline"
2089 },
2090 "widgets_values": [
2091 ""
2092 ],
2093 "widgets_values_named": {
2094 "value": ""
2095 },
2096 "color": "#223",
2097 "bgcolor": "#335"
2098 },
2099 {
2100 "id": 69,
2101 "type": "Note",
2102 "pos": [
2103 573.9302200846438,
2104 55.59540300266038
2105 ],
2106 "size": [
2107 314.4701158717141,
2108 88
2109 ],
2110 "flags": [],
2111 "order": 24,
2112 "mode": 4,
2113 "inputs": [],
2114 "outputs": [],
2115 "properties": [],
2116 "widgets_values": [
2117 "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."
2118 ],
2119 "widgets_values_named": {
2120 "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."
2121 },
2122 "color": "#432",
2123 "bgcolor": "#653"
2124 }
2125 ],
2126 "links": [
2127 [
2128 60,
2129 30,
2130 0,
2131 31,
2132 1,
2133 "SEEDVR2_DIT"
2134 ],
2135 [
2136 61,
2137 29,
2138 0,
2139 31,
2140 2,
2141 "SEEDVR2_VAE"
2142 ],
2143 [
2144 63,
2145 31,
2146 0,
2147 32,
2148 1,
2149 "IMAGE"
2150 ],
2151 [
2152 64,
2153 27,
2154 0,
2155 31,
2156 0,
2157 "IMAGE"
2158 ],
2159 [
2160 65,
2161 27,
2162 0,
2163 32,
2164 0,
2165 "IMAGE"
2166 ],
2167 [
2168 66,
2169 13,
2170 0,
2171 27,
2172 0,
2173 "IMAGE"
2174 ],
2175 [
2176 71,
2177 2,
2178 0,
2179 33,
2180 1,
2181 "CLIP"
2182 ],
2183 [
2184 83,
2185 31,
2186 0,
2187 43,
2188 0,
2189 "IMAGE"
2190 ],
2191 [
2192 84,
2193 1,
2194 0,
2195 33,
2196 0,
2197 "MODEL"
2198 ],
2199 [
2200 89,
2201 47,
2202 0,
2203 49,
2204 0,
2205 "LATENT"
2206 ],
2207 [
2208 90,
2209 3,
2210 0,
2211 49,
2212 1,
2213 "VAE"
2214 ],
2215 [
2216 96,
2217 50,
2218 0,
2219 51,
2220 0,
2221 "CONDITIONING"
2222 ],
2223 [
2224 98,
2225 51,
2226 0,
2227 47,
2228 2,
2229 "CONDITIONING"
2230 ],
2231 [
2232 99,
2233 50,
2234 0,
2235 47,
2236 1,
2237 "CONDITIONING"
2238 ],
2239 [
2240 100,
2241 49,
2242 0,
2243 13,
2244 0,
2245 "IMAGE"
2246 ],
2247 [
2248 138,
2249 50,
2250 1,
2251 47,
2252 3,
2253 "LATENT"
2254 ],
2255 [
2256 140,
2257 33,
2258 0,
2259 64,
2260 0,
2261 "MODEL"
2262 ],
2263 [
2264 141,
2265 64,
2266 0,
2267 47,
2268 0,
2269 "MODEL"
2270 ],
2271 [
2272 142,
2273 55,
2274 0,
2275 67,
2276 1,
2277 "STRING"
2278 ],
2279 [
2280 143,
2281 44,
2282 0,
2283 67,
2284 0,
2285 "STRING"
2286 ],
2287 [
2288 144,
2289 67,
2290 0,
2291 54,
2292 0,
2293 "STRING"
2294 ],
2295 [
2296 147,
2297 66,
2298 0,
2299 67,
2300 2,
2301 "BOOLEAN"
2302 ],
2303 [
2304 110,
2305 56,
2306 0,
2307 55,
2308 0,
2309 "CLIP"
2310 ],
2311 [
2312 127,
2313 0,
2314 0,
2315 55,
2316 1,
2317 "IMAGE"
2318 ],
2319 [
2320 112,
2321 57,
2322 0,
2323 55,
2324 4,
2325 "STRING"
2326 ],
2327 [
2328 149,
2329 68,
2330 0,
2331 55,
2332 5,
2333 "STRING"
2334 ],
2335 [
2336 129,
2337 35,
2338 0,
2339 55,
2340 6,
2341 "IMAGE"
2342 ],
2343 [
2344 130,
2345 36,
2346 0,
2347 55,
2348 7,
2349 "IMAGE"
2350 ],
2351 [
2352 131,
2353 34,
2354 0,
2355 55,
2356 8,
2357 "IMAGE"
2358 ],
2359 [
2360 132,
2361 37,
2362 0,
2363 55,
2364 9,
2365 "IMAGE"
2366 ],
2367 [
2368 133,
2369 38,
2370 0,
2371 55,
2372 10,
2373 "IMAGE"
2374 ],
2375 [
2376 134,
2377 39,
2378 0,
2379 55,
2380 11,
2381 "IMAGE"
2382 ],
2383 [
2384 135,
2385 40,
2386 0,
2387 55,
2388 12,
2389 "IMAGE"
2390 ],
2391 [
2392 128,
2393 14,
2394 0,
2395 55,
2396 15,
2397 "IMAGE"
2398 ],
2399 [
2400 95,
2401 33,
2402 1,
2403 50,
2404 0,
2405 "CLIP"
2406 ],
2407 [
2408 101,
2409 0,
2410 0,
2411 50,
2412 2,
2413 "IMAGE"
2414 ],
2415 [
2416 102,
2417 14,
2418 0,
2419 50,
2420 3,
2421 "IMAGE"
2422 ],
2423 [
2424 151,
2425 54,
2426 0,
2427 50,
2428 4,
2429 "STRING"
2430 ],
2431 [
2432 92,
2433 4,
2434 0,
2435 50,
2436 5,
2437 "INT"
2438 ],
2439 [
2440 93,
2441 4,
2442 1,
2443 50,
2444 6,
2445 "INT"
2446 ],
2447 [
2448 103,
2449 35,
2450 0,
2451 50,
2452 10,
2453 "IMAGE"
2454 ],
2455 [
2456 104,
2457 36,
2458 0,
2459 50,
2460 11,
2461 "IMAGE"
2462 ],
2463 [
2464 105,
2465 34,
2466 0,
2467 50,
2468 12,
2469 "IMAGE"
2470 ],
2471 [
2472 106,
2473 37,
2474 0,
2475 50,
2476 13,
2477 "IMAGE"
2478 ],
2479 [
2480 107,
2481 38,
2482 0,
2483 50,
2484 14,
2485 "IMAGE"
2486 ],
2487 [
2488 108,
2489 39,
2490 0,
2491 50,
2492 15,
2493 "IMAGE"
2494 ],
2495 [
2496 109,
2497 40,
2498 0,
2499 50,
2500 16,
2501 "IMAGE"
2502 ]
2503 ],
2504 "groups": [
2505 {
2506 "id": 1,
2507 "title": "SeedVR2, detail improvement/upscale (Optional)",
2508 "bounding": [
2509 2517.987556918573,
2510 300.8396022052916,
2511 1700.892061950401,
2512 720.891004666465
2513 ],
2514 "color": "#3f789e",
2515 "flags": []
2516 },
2517 {
2518 "id": 2,
2519 "title": "Load Models",
2520 "bounding": [
2521 -56.477069239445754,
2522 301.30881096732963,
2523 820.5269848739775,
2524 458.7669928969637
2525 ],
2526 "color": "#A88",
2527 "flags": []
2528 },
2529 {
2530 "id": 4,
2531 "title": "Gen settings",
2532 "bounding": [
2533 765.363324094742,
2534 300.3678327647891,
2535 955.6555488561293,
2536 521.2029016766687
2537 ],
2538 "color": "#3f789e",
2539 "flags": []
2540 },
2541 {
2542 "id": 6,
2543 "title": "Generation & Decode",
2544 "bounding": [
2545 1722.1035748955887,
2546 299.3191815184262,
2547 748.4996121415174,
2548 515.3618843956658
2549 ],
2550 "color": "#3f789e",
2551 "flags": []
2552 },
2553 {
2554 "id": 8,
2555 "title": "Ref Img 2",
2556 "bounding": [
2557 -51.276444655040876,
2558 1149.2926099300792,
2559 403.98016195908804,
2560 359.98557054352796
2561 ],
2562 "color": "#3f789e",
2563 "flags": []
2564 },
2565 {
2566 "id": 9,
2567 "title": "Ref Img 4",
2568 "bounding": [
2569 -50.346603249544785,
2570 1898.2093006739158,
2571 399.27417850158065,
2572 376.30600776809547
2573 ],
2574 "color": "#3f789e",
2575 "flags": []
2576 },
2577 {
2578 "id": 10,
2579 "title": "Ref Img 6",
2580 "bounding": [
2581 355.50713499613937,
2582 763.3068265994915,
2583 389.59024811997847,
2584 382.7374977196504
2585 ],
2586 "color": "#3f789e",
2587 "flags": []
2588 },
2589 {
2590 "id": 11,
2591 "title": "Ref Img 7",
2592 "bounding": [
2593 356.7063179168554,
2594 1152.1282377037212,
2595 396.4815129065377,
2596 358.15204880641636
2597 ],
2598 "color": "#3f789e",
2599 "flags": []
2600 },
2601 {
2602 "id": 12,
2603 "title": "Ref Img 9",
2604 "bounding": [
2605 354.59602350093303,
2606 1900.899229146382,
2607 390.7522708384979,
2608 387.6302249555231
2609 ],
2610 "color": "#3f789e",
2611 "flags": []
2612 },
2613 {
2614 "id": 13,
2615 "title": "Ref Img 8",
2616 "bounding": [
2617 357.85843456607927,
2618 1513.5017388117458,
2619 392.41956836884765,
2620 379.9614348793259
2621 ],
2622 "color": "#3f789e",
2623 "flags": []
2624 },
2625 {
2626 "id": 14,
2627 "title": "Ref Img 5",
2628 "bounding": [
2629 -51.30863574229829,
2630 2275.844055329408,
2631 399.674770544018,
2632 399.09877759641086
2633 ],
2634 "color": "#3f789e",
2635 "flags": []
2636 },
2637 {
2638 "id": 15,
2639 "title": "Ref Img 3",
2640 "bounding": [
2641 -50.42060574898747,
2642 1511.3982360875432,
2643 403.74490801336003,
2644 381.1901727313059
2645 ],
2646 "color": "#3f789e",
2647 "flags": []
2648 },
2649 {
2650 "id": 16,
2651 "title": "Ref Img 1",
2652 "bounding": [
2653 -52.34467073449487,
2654 760.6428792425585,
2655 404.92894800444094,
2656 389.33044766998967
2657 ],
2658 "color": "#3f789e",
2659 "flags": []
2660 },
2661 {
2662 "id": 17,
2663 "title": "Refiner",
2664 "bounding": [
2665 546.9631682502754,
2666 -432.0389298696163,
2667 1197.4217145350226,
2668 728.7799126130972
2669 ],
2670 "color": "#3f789e",
2671 "flags": []
2672 }
2673 ],
2674 "config": [],
2675 "extra": {
2676 "ds": {
2677 "scale": 0.43056764313424634,
2678 "offset": [
2679 868.0135472497877,
2680 1223.3497634546538
2681 ]
2682 },
2683 "frontendVersion": "1.58.5",
2684 "image_studio": {
2685 "release": "v21.0.0",
2686 "source_api": "examples/api/H3_REFERENCE_EDIT_API.json"
2687 },
2688 "VHS_latentpreview": false,
2689 "VHS_latentpreviewrate": 0,
2690 "VHS_MetadataImage": true,
2691 "VHS_KeepIntermediate": true,
2692 "zipTies": [
2693 {
2694 "id": 2,
2695 "pos": [
2696 1207.2808067492067,
2697 446.33818348783876
2698 ],
2699 "angle": -0.042109865618229124,
2700 "links": [
2701 95,
2702 101,
2703 90,
2704 102,
2705 106,
2706 103,
2707 104,
2708 105,
2709 107,
2710 108,
2711 109
2712 ],
2713 "color": "#d6409f"
2714 },
2715 {
2716 "id": 4,
2717 "pos": [
2718 975.9498791625452,
2719 -136.62717799413255
2720 ],
2721 "angle": -1.049885710796137,
2722 "links": [
2723 127,
2724 128,
2725 129,
2726 131,
2727 130,
2728 110,
2729 132,
2730 135,
2731 133,
2732 134
2733 ],
2734 "color": "#d6409f"
2735 },
2736 {
2737 "id": 5,
2738 "pos": [
2739 890.6806625113915,
2740 578.9739125243326
2741 ],
2742 "angle": -1.1173259514489047,
2743 "links": [
2744 127,
2745 95,
2746 128,
2747 129,
2748 130,
2749 131,
2750 110,
2751 90,
2752 101,
2753 132,
2754 135,
2755 102,
2756 133,
2757 134,
2758 103,
2759 106,
2760 104,
2761 105,
2762 107,
2763 109,
2764 108
2765 ],
2766 "color": "#f76b15"
2767 },
2768 {
2769 "id": 6,
2770 "pos": [
2771 802.5527094911258,
2772 1054.3925699056624
2773 ],
2774 "angle": -1.3668634370516246,
2775 "links": [
2776 128,
2777 129,
2778 130,
2779 131,
2780 135,
2781 102,
2782 133,
2783 134,
2784 103,
2785 104,
2786 105,
2787 107,
2788 109,
2789 108
2790 ],
2791 "color": "#d6409f"
2792 },
2793 {
2794 "id": 7,
2795 "pos": [
2796 704.7409742646253,
2797 1580.7898247540973
2798 ],
2799 "angle": -1.258439603000984,
2800 "links": [
2801 129,
2802 130,
2803 131,
2804 135,
2805 103,
2806 104,
2807 105,
2808 109
2809 ],
2810 "color": "#12a594"
2811 },
2812 {
2813 "id": 8,
2814 "pos": [
2815 753.8392081392327,
2816 1285.317477788427
2817 ],
2818 "angle": -1.2944537534263223,
2819 "links": [
2820 128,
2821 129,
2822 130,
2823 131,
2824 135,
2825 102,
2826 134,
2827 103,
2828 104,
2829 105,
2830 109,
2831 108
2832 ],
2833 "color": "#ededed"
2834 }
2835 ]
2836 },
2837 "version": 0.4
2838}