license: apache-2.0
library_name: diffusers
pipeline_tag: text-to-image
Not Ready
This is a WIP and not ready for use. This is an early testing version for research and development. You may know what this is and how to use it, if so, feel free, but it will change as I continue to develop it. I plan to do many updates to it frequently. So you may want to set a revision if you intend to use it anyway.
What is this?
FLUX.1-schnell is an amazing distilled model with an apache 2.0 license. However, it is not finetunable. LoRAs, IP adapters, control nets, etc, cannot be trained on it because it is distilled. The goal of this project is to finetune a non-distilled version of it that can be used as a training base to train adapters for FLUX.1-schnell.
Current Issues
Since we are breaking the distillation, this model will need many steps and guidance to produce good results. Currently, this model, like the schnell version, does not have guidance embeddings. Because of this (and possible other factors) images generated with this model will not look great. However, this hopefully will not affect training, since guidance is not used during training. The things trained on this model are intended to be used on the schnell version anyway. I am working on training guidance embeddings for it, but hopefully it will work as a training base without them.