Support schnell lora training

[phyllispeng123]

Hi @kohya-ss, I just find the current training script or code cannot be applied to FLUX schnell directly. According to @sdbds, it seems the shift problem? There are a lot of users prefer schnell, can you help to launch a schnell version for lora training ? that will be very helpful !!!!!!
I change the script to the following, but it still gets bad result.

--timestep_sampling uniform \
--discrete_flow_shift 1.0 \
--guidance_scale 0.0 \

[kohya-ss]

Since schnell is a model distilled for fast generation, I don't think it is suitable for fine tuning or LoRA training. So, it seems like one way is to train LoRA on dev and apply it to schnell.

[phyllispeng123]

Since schnell is a model distilled for fast generation, I don't think it is suitable for fine tuning or LoRA training. So, it seems like one way is to train LoRA on dev and apply it to schnell.

Thanks for your advice ~ I will have a try on dev first