lllyasviel
Here’s a better solution(?) to preloading models before generation.
Forge recently changed how models are loaded, which was a sensical thing to do as selecting a flux model will take forever to load in the background. My idea is simple.
If it’s possible, make it so the model only loads on “generate” for the first generation, then stores the loaded model for future generations until a different model is used to generate an image— at which point it loads that model instead.
This would speed up Forge significantly in my workflow!
I guess this is what destroyed performance for me, +20-30 sec per generation after update. I don't even care about flux.
I guess this is what destroyed performance for me, +20-30 sec per generation after update. I don't even care about flux.
Why?
I guess this is what destroyed performance for me, +20-30 sec per generation after update. I don't even care about flux.
Why?
it's not been finetuned to impress me enough for realism etc
it's only benefit in my eyes is the prompt adherence but its so slow for me that it's not worth it, I'll revisit it soon im sure once it's been looked at by others for a while
SDXL fits in my 6 gb vram with --unet/clip-in-fp8-e5m2, Flux does not. Also it's slower and not finetuned yet.