Neural image codecs have long struggled to balance pixel-wise accuracy with how natural the result looks. The new arXiv paper "PerQ: Inverse Generative Modeling for Neural Image Compression via Quantization Error Compensation" highlights this tension. Distortion-optimized codecs, the authors note, keep pixel fidelity but tend to blur at low bitrates. The abstract begins to contrast these with generative codecs before the text cuts off, leaving the full comparison incomplete.
To address this, the authors introduce PerQ, which they describe as inverse generative modeling. The method works by compensating for quantization error, the information lost when neural codecs compress an image. The abstract does not provide implementation details, but the title and framing indicate that PerQ targets quantization error as a key lever for improving reconstruction quality.
The paper is announced as a new arXiv submission (2610.27374v1). As with many early preprints, the abstract offers only a high-level view. The full method and experimental results would need to be read to assess how well PerQ balances the two objectives.