Sdcpp integration
Very nice work. I'm impressed by how similar the image quality is compared to the original Flux Dev. Surprisingly, I feel like prompt adherence is what's a bit degraded (especially with longer prompts) , but "no free lunch" I guess 😅.
I added support for this model to stable-diffusion.cpp via this PR. (You can also try it on my fork of sdcpp, wich has the support already merged in).
It runs nicely on 8GB GPU with q4_k quantization.
Thanks for the great work @stduhpf !!
Re: long prompt adherence (or lack thereof), the alpha version was trained on prompts shorter than 77 CLIP tokens, this may explain poor adherence to longer ones, but a new version is underway which is being trained on a larger and more diverse set of prompts including longer ones. Thanks again!
Support has been merged into master.
cuda, q4_k, 1024x1024 and diffusion-flash-attention uses 5.5gig vram during sampling.
and yea, text adherence is not great...