NOT ABLE TO LORA FINE TUNE ON SANA SPRINT MODELS

[pgagi-abhishek-khatri]

trafficstars

I want to LoRA fine-tune on the Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers, but I guess currently we can LoRA fine-tune on only the Efficient-Large-Model/Sana_1600M_1024px_BF16. I found the script here: https://github.com/NVlabs/Sana/blob/main/asset/docs/sana_lora_dreambooth.md

But when I replace the model with the sprint model, it produces some errors like invalid parameter passes 'batch size'. This code is not yet meant for sprint models. But then, is it impossible to have LoRA DreamBooth Fine Tuning on Sana Sprint Models?

Or am I doing something wrong?

[lawrence-cj]

SANA-Sprint is a distilled model with scm-loss (non diffusion loss). I'm not sure if the original diffusion loss is still work on it. I never have a try.

[nbardy]

Has anyone tried this? I feel like in theory training a LORA on the Sona Sprit weights with the standard DDIM should still work. But haven't had a chance to try.