LightChaser

I have tried to split the vae decoder's upsampling part. I confirm that the seams is from global-aware operators, specifically the attention, most of which can be safely removed except...

Hello, I have completed a wild hack that achieves exactly what you may want! https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111 This is an upscaler WITHOUT image post-processing. Everything is in latent space. The repo includes...

> > I have tried to split the vae decoder's upsampling part. I confirm that the seams is from global-aware operators, specifically the attention, most of which can be safely...

Hello I'm the original author of this engineering implementation. I'm willing to make a pr but I have little time，As I'm exploring the possibility of tiling unet (not folding it!)...

Yes it is very challenging. However, the text information is injected via the cross attention mechanism, where the (QK/d^-2) * V is a linear process. so you can just calc...

I wonder if we have the possibility to implement by ourselves with necessary hacking tricks? My project really need this feature and I do not how to achieve this (I'm...

I have done it with custom js in my extension. Please give me a star if you use that.

It lies in the only js file in my automatic1111 extension. I didn't rely on gradio to paint it.