TL;DR
Learned Image Compression (LIC) models often suffer from "multi-generation loss," where image quality rapidly degrades after repeated cycles of encoding and decoding (e.g., during editing or transcoding). This paper identifies quantization drift as a primary culprit and proposes three simple yet effective fixes—straight quantization, a reversibility loss function, and channel relaxation—allowing AI codecs to remain stable even after 50 consecutive re-encodings.
What Problem Does This Paper Address?
In the real world, images aren't just compressed once. They are edited, uploaded, downloaded, and transcoded across different platforms. This process is called Successive Image Compression (SIC).
While traditional codecs like JPEG or BPG are relatively stable over multiple generations, modern AI-based Learned Image Compression (LIC) models are notoriously fragile. If you compress an image with a standard LIC model, decode it, and then compress that result again, the quality often "death spirals." After a few dozen rounds, the image becomes plagued by blurring, strange color casts, and high-frequency noise. This instability has been a major barrier to adopting AI compression in professional workflows like video editing or social media pipelines.
Methods
The authors pinpointed the source of the "death spiral" and proposed a three-pronged solution:
- Straight Quantization (SQ): Most LIC models use "corrected quantization," where the model subtracts a predicted mean ($\mu$) from the latent representation before rounding. Because $\mu$ is calculated based on surrounding pixels (context), a tiny error in one pixel causes a "chain reaction" that shifts the means of all subsequent pixels. The authors suggest switching to Straight Quantization—rounding the values directly—and using a Straight-Through Estimator (STE) during training to bridge the gap between discrete rounding and continuous gradients.
- Reversibility Loss (RL): Standard training only cares about the final reconstructed image. The authors added a new constraint to the loss function that forces the mathematical transform (the "encoder" and "decoder" parts) to be more reversible. Specifically, they penalize the difference between the original image and a version passed through the encoder-decoder pipeline without the rounding step.
- Channel Relaxation (CR): The authors found that "bottlenecking" the data into too few channels at the latent layer forces the model to lose information during the transform stage. By increasing the number of channels ($M$) in the latent representation, they "relax" this bottleneck, allowing the transform to be more stable across generations.
Datasets
The models were trained using the Flicker 2W dataset (20,745 high-quality images). For evaluation and benchmarking of the multi-generation robustness, the authors used three standard industry datasets:
- Kodak (24 images)
- CLIC Professional
- Tecnick
Key Insights and Ideas
- The "Chain Reaction" of Context: The most surprising insight is that the very feature that makes AI codecs efficient—using context to predict pixel values—is what makes them unstable. By making the quantization of one pixel dependent on the decoded values of others, any small error (quantization noise) is amplified and propagated through the entire image in the next generation.
- Variational Models vs. INNs: The researchers tested Invertible Neural Networks (INNs), which are mathematically designed to be reversible. Surprisingly, INNs performed worse in SIC tasks. Because they are perfectly reversible, they pass high-frequency "unfriendly" noise directly into the latent domain, where it accumulates. Standard variational models act like low-pass filters, which actually helps them adapt to errors better than INNs.
- Stability without Architecture Changes: Perhaps the most practical insight is that these improvements don't require redesigning the neural network. They are achieved through changes in the training loss and the quantization logic used during inference.
Key Results
The proposed methods significantly outperformed previous attempts to fix multi-generation loss (like Feature Identity loss).
- Stability: Using the "SQ+RL" or "SQ+CR" methods, the image quality degradation converges (stops getting worse) very quickly.
- Long-term Performance: After 50 cycles of re-encoding, the average PSNR (a measure of image quality) reduction was less than 1.0 dB.
- Competitive with BPG: Even after 50 rounds of re-encoding, the AI model maintained a rate-distortion performance comparable to the first compression of the traditional BPG codec.
Limitations and Future Work
- Model Complexity: The "Channel Relaxation" (CR) method improves robustness but increases the number of parameters in the model, which might be a drawback for mobile or edge devices.
- Rate-Distortion Trade-off: The "Reversibility Loss" (RL) slightly impacts the initial compression efficiency (the first-time encode) to gain long-term stability. Finding a way to get "perfect" stability with zero hit to the initial bitrate remains an open challenge.
- Cross-Codec Robustness: This paper focuses on "intra-model" robustness (re-encoding with the same model). Future work could explore "inter-model" robustness, where an image is passed between different AI codecs or between an AI codec and a traditional one like JPEG.
Summary updated