ByteDance

/

Hyper-SD

like 880

Hi, @KeMiliUs
Thanks for your attention!
The answer is YES and we're working on it.
It will be released ASAP.

That's really exciting! @Yanzuo , how long does the distillation normally take? In your paper, you mentioned 200 hours on A100.

@radames
This is not as simple as using the same method and changing the model.
SD3 uses a completely different way of modeling the scheduler so it might takes more time for exploration to get better results.
Sorry that we can't give an exact time.

Hi @Yanzuo , thanks for the information, and no worries. I was just wondering about the time requirements.
Thanks for doing it. Hyper SD has been my favorite model to play with. I'm looking forward to playing with Hyper-SD3.

Will it correct the strange human body deformities?

Hi @zhaoqi , we'll try best.

mark! I really like your model!

any updates on this?

any updates on this?

@akiokawahito @cpanay @Sumail @radames
Sorry for taking such a long time.
SD3-related LoRAs are released now.

Any plan for Flux? SD3 seems abandoned by the community, everyone is talking about Flux.

@akiokawahito
Absolutely. We have noticed this issue so this SD3 LoRA might be just for academic purpose.
We have posted our plan about FLUX in the NEWS of model card.
It will certainly be coming next week : )

@akiokawahito @cpanay @Sumail @radames
Sorry for taking such a long time.
SD3-related LoRAs are released now.

good news, wish to see your FLUX LoRA later.
加油！

@akiokawahito @Sumail
8-steps and 16-steps LoRAs for FLUX.1-dev are available now.
Lower step would be coming soon.

is this setting ok?

@akiokawahito
We haven't tested on ComfyUI. But the results seem to be okay : )
I'm not sure if the lora scale set in diffusers is the same as in ComfyUI.

Cool, everything is OK!