tfernd / sd-fused Goto Github PK

Some useful references:
https://github.com/huggingface/diffusers/tree/main/examples/dreambooth
https://github.com/bmaltais/kohya_ss/blob/master/train_db_fixed_v6.py

Great job! Looking forward to using your code to do various custom training on top of the original sd model~

Very interested in the work you're doing. Speed and memory efficiency are crucial for anyone trying to generate at scale.

We've implemented Token Merging: facebookresearch/ToMe#7

This has a speed overhead increase of about 15% from the naive implementation at 512x512, although this goes up as array sizes increase. The memory overhead reduction is significant, and can allow for much larger image generation.

Can you implemente clip guided stable diffusion? like this,https://github.com/huggingface/diffusers/blob/main/examples/community/clip_guided_stable_diffusion.py,I have tried, using one and more to generate a good picture quality, it is worth joining!
Looking forward to your masterpiece!

Hello,The recent use of the effect has become worse, is there some adjustments that have been made to affect the generation effect?

Thanks for this library! It works just fine on my RX6700XT with ROCM PyTorch installed, and with ToMe enabled it gives speeds similar to other implementations (that don't have it).

I wonder if it's feasible to try to add cross attention like in https://github.com/Doggettx/stable-diffusion (I think it's different from the "original" split attention)? Because with just that optimization with auto's webui (which doesn't really have any other optimizations for my GPU) I'm getting the same speed as with your code. It's also implemented in auto's web ui but that repo doesn't have a license, so using code from it wouldn't really be possible.

I've also checked Birch-san's SD fork with ToMe implemented, and it also gives similar speeds to having this Doggettx's split attention, so maybe it's possible to combine the two.

I know this isn't really relevant for modern NVIDIA GPUs as they have xformers, but I think it'd still be nice for older NVIDIA GPUs and AMD GPUs :)

Hello, you have a great library! I've tried it many times and the results are awesome! But the shortcoming is that it lacks the correction of face details and the restoration of image details. Can you consider adding post-processing algorithms such as RealESRGAN, ESRGAN, CodeFormer, GFPGAN, etc.? Looking forward to your improvement, thanks!