Lower bitrate experiment

[bblgbr]

Hello,

I greatly admire the work done in DiffEIC, which is based on diffusion models. I am interested in fine-tuning a model to achieve a lower bpp, such as 0.01. To do this, do I simply need to adjust the l_bpp_weight parameter in the DiffEIC/configs/model/diffeic.yaml file to a value lower than 16? Have you conducted any experiments with such low bpp, and would you be able to recommend an appropriate value for this parameter or share a checkpoint?

Thank you for your help!

[huai-chang]

Hi, thank you for your kind words and for your interest in our work. Unfortunately, we have not experimented with bpp values as low as 0.01. If you would like to train a model targeting such a low bpp, you can increase the l_bpp_weight parameter in DiffEIC/configs/model/diffeic.yaml to larger values, such as 32 or 64. However, please note that the reconstruction quality will likely degrade significantly under such extreme compression.

[huai-chang]

You may consider continuing training based on the 1_2_1 checkpoint, which may reduce training time.

[bblgbr]

Thank you for your response. I have fine-tuned with l_bpp_weight={19, 24, 32, 64}, and in our scenario, the image quality indeed decreased significantly. We will report the corresponding results in our paper and cite your work. Once again, thank you for your contributions to the compression community.