martyn / safetensors-merge-supermario

Merge safetensor files using the technique described in "Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Inference step

cuishuhao opened this issue · comments

Hello,
thanks for sharing!
I wonder how you get the merged image? Inference for only One step or 50steps? It will be better to provide the prompt and seed!
Thanks

Thanks! The linked json has the exact ComfyUI settings but the steps were set to 10 for all the images using the SDTurboScheduler. Single step doesn't work on the merged model. I didn't try to tune the lambda or p values to improve it.