FlowCodec: One-step image compression at <0.05 bpp using AI priors

FlowCodec introduces a streamlined approach to generative image compression by leveraging large-scale pretrained text-to-image models as priors. The framework decomposes the pipeline into two stages: Latent Compression, which encodes clean latents into bitrate-constrained noisy latents, and Latent Transport, which uses a single-step flow to refine these noisy latents back toward clean ones. The key innovation is that FlowCodec requires no additional conditioning signals or auxiliary branches, dramatically simplifying integration compared to prior diffusion-based compression methods.

Extensive experiments demonstrate that FlowCodec preserves high visual quality even at bitrates below 0.05 bits per pixel (bpp) by accessing two powerful pretrained backbones: Qwen-image-2512 and FLUX.1-dev. The Qwen variant particularly excels in perceptual metrics like LPIPS and DISTS, while both variants achieve higher PSNR and significantly faster encoding than existing one-step diffusion codecs. Impressively, FlowCodec keeps trainable parameters under 0.54% of the generative backbone, enabling flexible multi-bitrate support with minimal overhead. This design suggests a practical path to generalizing large-scale generation models for compression without costly retraining or engineering.