Spaces:
Running on Zero
Running on Zero
add 4 step turbo lora
#2
by akhaliq HF Staff - opened
Adds the 4-step distilled LoRA from akhaliq/MOSS-Image-SFT-4-steps as a turbo model choice, keeping Base/SFT as defaults. The fp32 LoRA wraps the SFT transformer at startup (unmerged) and is toggled per request by alpha, no third pipeline in memory. Turbo runs CFG 1 with resolution-mapped steps (4/6/8 per the model card). New file turbo_lora.py is vendored from the model repo with attribution.
akhaliq changed pull request status to open