Regarding quantization parameter adjustment

[QQQ27]

Excellent work! Last year, I used the deepcompressor library to quantize a flux 8b model. The quantized model showed excellent image quality and maintained good consistency before and after quantization.
Now, I'm planning to use the deepcompressor library to quantize my own model, which is based on flux 8b and uses the default quantization settings of flux.1-dev. However, the quantization results are not ideal. The SSIM evaluation metric on the MJHQ dataset is only 0.72 for the quantized model compared to the pre-quantized model. The image quality of the quantized model is as follows, the image is very blurry:

Does it seem that the quantization accuracy is insufficient? Can adjusting the quantization parameters improve this problem?

[Ahmaddimran]

@QQQ27 Hi, I wanted to know how you are doing the inference. I am having issue in loading model's weight in particularly when adding the branch. Can you elaborate on this please