How to train a lora with distilled flux model, such as flux-schnell???

[Johnson-yue]

Is your feature request related to a problem? Please describe.
I can use flux as base model to train a lora, but it need 20 steps , it cost a lot of time , and I want to train a lora base on distill model to implement use fewer step make a better image, such as based on flux-schnell model train a lora it only need 4 steps can generate a good image !! and I can train many lora like this, only need 4 steps generated

Describe the solution you'd like.
I need a script , maybe it locate in examples\dreambooth\train_dreambooth_lora_flux_schennl.py
I want to know to train a lora based on distilled model and get a good result ?

Describe alternatives you've considered.
I want to train many lora for base model( flux or flux-schnell), not only one lora , and I want to generated with fewer steps. So , I want to train loras with distilled model ... how to implment it ? I test scripts : train_dreambooth_lora_flux.py by modify based mode from flux to flux-schnell ,but the result is bad...

Additional context.
any other implement method is OK ,

[dxqb]

you generally don't. There is a training adapter here https://huggingface.co/ostris/FLUX.1-schnell-training-adapter but I'd consider this expert use and not something that diffusers as a generic library will want to provide a script for (correct me if I'm wrong, I'm not associated with diffusers)