Open geroldmeisinger opened 1 month ago
Sorry, we won't have the bandwidth to work on that script right now. I will leave it open to the community in case someone from the community wants to pick it up.
Cc: @kashif as well since he worked on the SD Cascade scripts.
Hi @sayakpaul, could you pls tell how may I start working on this? Like how may I proceed in doing such tasks? Any help will be appreciated. Thanks
Is your feature request related to a problem? Please describe. Last year I wrote a long article on how to train controlnets using diffusers, and trained two controlnets using diffusers. A few months ago Stable Cascade was released which requires less resources than SDXL and also should make controlnet training for high quality diffusion models more viable. I tried to run the official training script (batchsize=4, bf16) but ran into OOM. I'd hope that a diffusers implementation would provide more optimization and lower VRAM requirements.
Describe the solution you'd like. Please provide a
examples/controlnet/train_controlnet_stablecascade.py
Describe alternatives you've considered. The official training script => OOM
Additional context. A training script for the Würstchen architecture was already anticipated a long time ago (see https://github.com/huggingface/diffusers/issues/5071).
Pretty please!