voltaML-fast-stable-diffusion

Fooocus MRE for SDXL. The quality is just so much better than a1111 or vlad's fork. And its stable, with a good UI: https://github.com/MoonRide303/Fooocus-MRE/commits/moonride-... I use VoltaML for SD 1.5: https://github.com/VoltaML/voltaML-fast-stable-diffusion/tre... Next on my todo list is trying ComfyUI with AIT: https://github.com/FizzleDorf/ComfyUI-AIT

Also, VoltaML has a good reference GPU AITemplate SD 1.5 implementation: https://github.com/VoltaML/voltaML-fast-stable-diffusion/tre... The speed jump is massive on my desktop GPU, probably even more dramatic on cloud hardware, and it may support some things (weight swapping/lora swapping/resolution changing/controlnet) better than JAX.

This is the best I've tried so far, but no mac support I don't think. Its a feature packed fork of Fooocus, which was developed by the orginal ControlNet dev. The quality you can get from small prompts is mind boggling: https://github.com/MoonRide303/Fooocus-MRE For base SD 1.5, I use Volta, because its fast: https://github.com/VoltaML/voltaML-fast-stable-diffusion/com... Really good SD 1.5 image quality comes from gratuitous use of finetunes, LORAs, controlnet and other augmentations. So you can, say, trace a base image for structure, specify prompting in certain areas of the image and so on. InvokeAI is actually quite feature packed, and has lots of these augmentations hidden in the nodes UI, but Volta and other UIs also expose them more directly.

> how many iterations are usual these days? I dunno about a precise number...Less than 30? The UniPC scheduler in particular can get great results below 20 iterations. Karras DPM schedulers are pretty good with 30. > In terms of unmodified diffusers, you mean fine tuned to whatever they are working on? No, people hack non-default augmentations onto the vanilla diffusers (or SAI) pipe to increase speed or increase quality. Some examples included the channels last memory format, downsampling some stages to FP16, TomeSD, self attention, and some more user-guided augmentations like controlnet or tiled diffusion. "alternate" runtimes like the ones available via torch.compile or AITemplate provide some pretty massive speed gains (~2x) if you need them, but are less common since they are hard to install and HF diffusers only (where most of the S

Llama.cpp/ggml is uniquely suited to llms. The memory requirements are huge, quantization is effective, and token generation is surprisingly serial and bandwidth bound, making it good for CPUs, and an even better fit for ggml's unique pipelined CPU/GPU inference. ...But Stable Diffusion is not the same. It doesn't quantize as well, the unet is very compute intense, and batched image generation is effective and useful to single users. Its a better fit for GPUs/IGPs. Additionally, it massively benefits from the hackability of the Python implementations. I think ML compilation to executables is the way for SD. AITemplate is already blazing fast [1], and TVM Vulkan is very promising if anyone will actually flesh out the demo implementation [2]. And they preserve most of the hackability of the pure PyTorch implementations. 1: https://github.co

This kind of drama is just sad. I dont know if you are OP, but plenty of other UIs have interrogator code, like https://github.com/VoltaML/voltaML-fast-stable-diffusion/tre...

is it possible to add volta or xformers for a massive speed increase? https://github.com/VoltaML/voltaML-fast-stable-diffusion