Stable diffusion generated samples using quantized checkpoint looks strange

[colorjam]

Thanks for the contribution of the author.

I follow the inference script to run the txt2img demo:

python scripts/txt2img.py --prompt "a puppy wearing a hat" --plms --cond --ptq --weight_bit 8 --quant_mode qdiff --no_grad_ckpt --split --n_samples 5 --resume --outdir ./experiment/weight_only_w8 --cali_ckpt ./models/sd_w8a8_ckpt.pth --ckpt ../models/sd/sd-v1-4.ckpt

generated samples:

[list0830]

I have met the same problem

[xuchen-dev]

你们跑了多久 这个跑的好慢