Type | |
Stats | 257 0 |
Reviews | (33) |
Published | Apr 24, 2025 |
Base Model | |
Hash | AutoV2 7CCB46E680 |
I will not be locking this model behind an early access fee, but buzz donations are greatly appreciated. Or, just post your generations to this page!
UPDATE: v0.2 is a set of FP8 models that can reliably render nude women and shirtless men. I've figured out some better training techniques for HiDream, and I've set it up to preserve as much as HiDream's incredible style variety as possible. There are fewer nudes in the example images this time around to demonstrate that this deserves to be your base model. :)
It can do:
Photos
Many art styles
Boobies
Vaginas
Pubic hair, sometimes.
All of the other stuff base HiDream can do.
It can't (yet) do:
Dicks
Sex acts
A number of different poses
Please note that I used comfyui-dynamic-prompts for prompt generation, so the crazy long prompts you're seeing here on civit aren't actually the prompts being sent to the encoder.
To really go all-in on art styles the way I did, you need my comfyui custom nodes. I also very strongly recommend merging my pull request (see READ THIS FOR BEST RESULTS below).
Original text:
v0.1 is a (partially) set of FP8 models (full, dev, and fast) that can reliably render topless characters (men and women, although HiDream renders shirtless men just fine on its own). This version doesn't yet render full nudity very well, but I'm working on that.
I'm working on expanding this in order of increasing difficulty. Next up are lady parts, then man parts (penises are hard. to train.), then possibly sex acts. v0.2 is still training, but the results are very promising.
READ THIS FOR BEST RESULTS:
CLIP and t5 don't do HiDreams any favors (I have a strong suspicion that t5 actively sabotages NSFW gens), particularly when rendering NSFW images. Update your ComfyUI to the latest master and load ONLY llama as your clip encoder (set the pulldown to hidream). If you're technically inclined, merge my pull request, which replaces the tensors full of zeros it sends as CLIP and t5 encodings with cached encoded empty prompts, as the zeroed tensors make HiDream a bit janky and unstable.