文件清单 线路:原始(huggingface.co) · 切换
下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。
说明
Woodland Camo BDU, H3 Image (and video?)
作者:andyx1976
Woodland Camo BDU, now for H3 Image. It also seems to work with H3 video and (much less well!) p-h3 image in my tests. But it was trained on H3 image, so i don't promise anything else :) But the little video had no start image, the uniforms are purely done by the lora. H3, left to it's own devices makes as "BDU a modern one, with the diagonal pockets and so on. You really want to use the trigger word here to overcome that, more than in other versions. Trigger word is, as always "Woodland Camo BDU" p-h3 worked kind of but details like prompted nametags are messed up. If is disable the p-h3 lora and run just a H3 video model as image model, the woodland works again. Not sure if there is a point to that over the image model.
The model can't keep up in complexity and perfect anatomy with Krea2, Kroma or Boogu. But hey, this resolution, at that speed is just a killer argument to end all arguments. And it works perfectly well with H3 finetunes like 10eros.
The lora, as always, is NOT trained on any NSFW content, but if you give the image/video model some wiggle room (lora at 0.8 or so) it can work with it quite well. However my dataset on a vintage uniform is mostly vintage images, so they are NOT at 4 Mpx, let alone 16. And i don't want to add too many ai generated images to fill the resolution gaps. So don't expect miracles. At least from me. The model seems to fill the detail gaps ok, but 16mpx is a bit of a quality stretch. The ones i use are btw Boogu Image ones, upscaled/refined with seedvr or Z-Image. It is the best working lora version imho in visual quality.
May also do another version later. This is already my second. AITookit defaults to rank 16 for H3 training. That is not enough. Even at 4000steps h3 takes a takes a rank16 lora as a mild suggestion. Rank32 works better.
