LORA training on flux1-dev-bnb-nf4-v2?

[iqddd]

Is that possible?
LORA trained on fp16 unet + fp16 T5XXL give poor generation results on bnb-nf4 (less character recognition, weaker style adherence, etc).
Web GUI from bmaltais/kohya_ss throws an error when starting training.

[iqddd]

Perhaps it is necessary to reverse conversion from NF4 to FP8/FP16?

[lllyasviel]

#1038

See also the section about lora precision

[iqddd]

Nothing about flux1-dev-bnb-nf4-v2-based training, just inference and related.