Summary
In this paper, the authors propose a novel architecture of lossy image compression using conditional diffusion models. They use an encoder to map images into contextual latent variables and train a conditional diffusion model under the guidance of these contextual latent variables as the decoder. Moreover, they add perceptual loss into the loss of diffusion model, so this approach can steer the tradeoff between distortion, rate and perceptual quality by changing the parameter of its loss and diffusion model. Finally, they conduct sufficient experiments to explore how the parameters influence the final performance and to compare with other approaches.
Strengths
1. The architecture based on the conditional diffusion model is interesting and creative, and as is shown from the experiments in 3.2, this architecture can also finish the reversed progress of the conditional diffusion model in a handful of steps.
2. Sufficient experiments have been conducted, including comparison with other methods in sixteen metrics, and exploration of how parameters like ρ, γ, and decoding steps influence the final performance.
2. Overall the results prove that this method can steer the tradeoff between distortion, rate and perceptual quality, and the final performance with specific parameters in the paper is also quite competitive in different metrics.
Weaknesses
1. Some implementation details are not illustrated clearly, including how to get the "17 steps" of the reversed progress in 4.1 which may be helpful to understand the schedule of the reversed progress, the cost of encoding and decoding time, and I think a picture including the image of every single step in 17 steps when predicting x0, may help to prove that the architecture can reconstruct the "texture" structure of the image in a handful of steps.
2. More qualitative and quantitative results may be helpful to prove the performance and practicability of the method. including images with different resolutions and different styles of images.
2. I think, when testing, the parameters of the diffusion model should be given more specifically, including the total number of steps, the true number of steps after resampling, and the type of schedule(cosine or linear), because these parameters can influence the decoding steps metioned in the paper.
3. In formula (7), I think it can only represent the situation of γ=0, that is the DDIM, but when γ>0, I think there should be an extra random noise according to the formula(predicting noise is changed to predicting x0)
$$
\boldsymbol{x}_{t-1}=\sqrt{\alpha_{t-1}} \underbrace{\left(\frac{\boldsymbol{x}_t-\sqrt{1-\alpha_t} \epsilon_\theta^{(t)}\left(\boldsymbol{x}_t\right)}{\sqrt{\alpha_t}}\right)}_{\text {"predicted } \boldsymbol{x}_0 "}+\underbrace{\sqrt{1-\alpha_{t-1}-\sigma_t^2} \cdot \epsilon_\theta^{(t)}\left(\boldsymbol{x}_t\right)}_{\text {"direction pointing to } \boldsymbol{x}_t \text { " }}+\underbrace{\sigma_t \epsilon_t}_{\text {random noise }}
$$
Questions
1. In section 4, I wonder why the images of the Kodak dataset are not resized while the others are resized, and how the diffusion model trained on an image size of 256×256 can be applied to images of different sizes when testing.
Rating
6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.
Confidence
3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
Limitations
Yes, the authors have considered the limitations and societal impact of neural compression, particularly when prioritizing perceptual quality, is the risk of misrepresenting data in the low bitrate regime.