Thank you!
This model is incredible.
I created a custom node to easily use it in ComfyUI:
https://github.com/easygoing0114/ComfyUI-uit-hidream-o1
Thank you for releasing it under the MIT license!
if they'd fix the blurry fuzziness (works well with your example image though :), it would be a decent image model with good reference image handling.
Edit: ai toolkit now supports it for lora training and man is it training fast. So hopefully people will fix it up.
I’ve created a fine-tuned model adjusted to produce clearer outputs. Since it was tuned using anime illustrations, it may result in slightly stronger contrast when used for photorealistic images.
Since there is no VAE post-processing applied, I think the fact that it’s easy to make further adjustments even on the base model is a strong point that many will appreciate.
i gotta say, i only do loras but it trains extremely fast. Ai Toolkit default settings it does 1.4 or so it/sec on my card which is fast than any other model i ever tried.
so i hope the same. And it learns clothing and stuff very quickly in the training. But skin is still always pure blur. for that we need a lore/fine tune that does more details not less.
Which is all a shame because there are good things too. Anatomy for example. After the absolute Anatomy CF that was ernie and even the mess Klein 9b and 4b make, hidream o1 very rarely gets anatomy wrong.
with the z-image upscale from the other thread, the blur issue is basically solved and the model actually useful. Even at .15 denoising it helps a lot, without destroying identity or details. .
But my loras don't work with the checkpoint- versions of the model. Aitoolkit does it in a way that is not compatible. It works with the dedicated hO1 nodes though.