Came to ask, can you do INT8 versions for the new turbo and aesthetic anima models that just came out? 🙏
Title. It would great if you could do Int8 Quantized like you did for base, to save some ram
I have uploaded int8convrot version of aesthetic-v1.0, aesthetic-v1.0b and turbo-v1.0.
- anima-aesthetic-v1.0-int8convrot.safetensors
- anima-aesthetic-v1.0b-int8convrot.safetensors
- anima-turbo-v1.0-int8convrot.safetensors
You can also quantize own quants model using comfy-dit-quantizer.
The command for int8convrot is as below.
python quantize.py configs/anima-int8rowwise.json models/anima-aesthetic-v1.0.safetensors models/anima-aesthetic-v1.0-int8convrot.safetensors --convrot -m percentile
(before running, you should activate venv of ComfyUI.)
Or you can use official(?) tool like,
Thank you, sorry I didn't know much about how to quantize them myself, also wasn't sure if it would work the same way for the newer finetunes.
edit: btw, I think Circlestone Labs plans to release aesthetic 1.1 soon (they said it was gonna be ready for like yesterday) to fix and reduce the noise and make it have smoother lines, in case you want to quantize that one too in the future
Yes, I will definitely quantize that.
By the way, There's also ComfyUI custom node to quantize model, so you can do it easily. (mine keeps more weights for quality, tho)
If you need to quantize some finetuned model, try it.
https://github.com/Starnodes2024/comfyui-starnodes-modelconverter