Improving Multi-generation Robustness of Learned Image Compression

Litian Li, Zheng Yang, Ronggang Wang

Abstract

Benefit from flexible network designs and end-to-end joint optimization approach, learned image compression (LIC) has demonstrated excellent coding performance and practical feasibility in recent years. However, existing compression models suffer from serious multi-generation loss, which always occurs during image editing and transcoding. During the process of repeatedly encoding and decoding, the quality of the image will rapidly degrade, resulting in various types of distortion, which significantly limits the practical application of LIC. In this paper, a thorough analysis is carried out to determine the source of generative loss in successive image compression (SIC). We point out and solve the quantization drift problem that affects SIC, reversibility loss function as well as channel relaxation method are proposed to further reduce the generation loss. Experiments show that by using our proposed solutions, LIC can achieve comparable performance to the first compression of BPG even after 50 times reencoding without any change of the network structure.

Introduction

In real-world scenarios, images undergo multiple cycles of encoding and decoding, a process known as Successive Image Compression (SIC). This occurs during social media uploads, cloud storage, and professional editing workflows. While traditional codecs like JPEG and BPG maintain relative stability across generations, modern AI-based Learned Image Compression (LIC) models are notoriously fragile. Standard LIC models often enter a "death spiral" where image quality rapidly degrades into blurring, color shifts, and noise after only a few dozen re-encodings. This instability significantly hinders the adoption of AI compression in practical pipelines.

Methodology

The researchers identify quantization drift as the primary cause of multi-generation loss. In context-based LIC models, the quantization of a pixel depends on a predicted mean ($\mu$) derived from surrounding decoded pixels. Small errors in one pixel propagate through this context, causing a chain reaction that shifts the means of subsequent pixels in future generations. To address this, three solutions are proposed:

  1. Straight Quantization (SQ): By switching to direct rounding of latent representations and using a Straight-Through Estimator (STE) during training, the model avoids the error-prone dependency on predicted means.
  2. Reversibility Loss (RL): A new constraint is added to the loss function to ensure the encoder-decoder transform is more mathematically reversible, penalizing differences between the original and reconstructed images even before rounding.
  3. Channel Relaxation (CR): Increasing the number of channels ($M$) in the latent representation "relaxes" the information bottleneck, allowing the transform to preserve more data across generations.

Experimental Results

The proposed methods were evaluated using the Flicker 2W dataset for training and standard benchmarks (Kodak, CLIC Professional, Tecknick) for testing.

Visual comparison and RD performance after 50 re-encoding cycles. (a) Visual results for various codecs on "kodim24.png". (b) Rate-Distortion performance showing the stability of the proposed SQ+RL and SQ+CR methods compared to the rapid degradation of baseline LIC (mbt18).

As shown in the figure above, the proposed "SQ+RL" and "SQ+CR" methods demonstrate remarkable stability. While the baseline mbt18 model suffers from severe artifacts and a massive drop in PSNR, the proposed solutions converge quickly. After 50 cycles, the PSNR reduction is less than 1.0 dB, maintaining performance comparable to the initial compression of the traditional BPG codec.

Discussion and Insights

A key finding is that the efficiency-boosting context prediction in AI codecs is precisely what causes instability. Interestingly, the study also found that Invertible Neural Networks (INNs) performed worse than standard variational models in SIC tasks. Because INNs are perfectly reversible, they tend to accumulate high-frequency noise, whereas variational models act as low-pass filters that help mitigate error propagation.

Conclusion

This work demonstrates that multi-generation robustness in LIC can be achieved without architectural changes. By modifying the quantization logic and training loss, AI codecs can match the stability of traditional standards while retaining their superior coding efficiency. Future work will focus on optimizing the rate-distortion trade-off introduced by reversibility constraints and exploring robustness across different heterogeneous codecs.

Summary updated

Abstract

Benefit from flexible network designs and end-to-end joint optimization approach, learned image compression (LIC) has demonstrated excellent coding performance and practical feasibility in recent years. However, existing compression models suffer from serious multi-generation loss, which always occurs during image editing and transcoding. During the process of repeatedly encoding and decoding, the quality of the image will rapidly degrade, resulting in various types of distortion, which significantly limits the practical application of LIC. In this paper, a thorough analysis is carried out to determine the source of generative loss in successive image compression (SIC). We point out and solve the quantization drift problem that affects SIC, reversibility loss function as well as channel relaxation method are proposed to further reduce the generation loss. Experiments show that by using our proposed solutions, LIC can achieve comparable performance to the first compression of BPG even after 50 times reencoding without any change of the network structure.
BibTeX
@inproceedings{li2023improving,
  author = {Litian Li and Zheng Yang and Ronggang Wang},
  title = {{Improving Multi-generation Robustness of Learned Image Compression}},
  booktitle = {ICME},
  year = {2023},
  url = {https://openreview.net/forum?id=lHMHeuzLp2},
}
Figure