Research map / Image generation / Image restoration
Correcting Diffusion-Based Perceptual Image Compression with Privileged End-to-End Decoder
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CorrDiff, a diffusion-based perceptual image compression method where an end-to-end decoder (ELIC-based) supplies a correction to the diffusion score function, with only per-step linear factors (128 bits total) sent as side information.
Problem
Diffusion-based image compression produces perceptually good images but struggles to guarantee distortion fidelity; integrating diffusion models with learned compression remains under-explored.
Approach
Theoretically derives a correction term for the score function using the original image visible at the encoder, approximates it with an end-to-end decoder's output and a per-timestep linear combination factor gamma_t chosen by minimizing a perceptual metric, and transmits only these few float factors alongside the quantized latent.
Key results
- Across 9 metrics on CLIC professional, CorrDiff achieves general superiority on perceptual metrics while also improving distortion over HiFiC, ILLM and CDC.
- Ablation BD rates on CLIC professional: Only End-to-End gives +22.177% perception BD rate; Only Diffusion gives +28.756% distortion BD rate; Direct score network gives +11.235% total.
- Correction factors are sent as float16 values, 8 DDIM steps, totaling 128 bits of side information.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use only the end-to-end decoder, removing diffusion correction (ablation) | CorrDiff · CLIC professional | Average BD rate across perception and distortion: Perception: +22.177; Distortion: -3.923; Total: +9.127 | Mixed |
| Use only diffusion, removing the end-to-end decoder correction (ablation) | CorrDiff · CLIC professional | Average BD rate across perception and distortion: Perception: +4.833; Distortion: +28.756; Total: +16.794 | Hurt |
| Use the score network for one-step direct reconstruction (ablation) | CorrDiff · CLIC professional | Average BD rate across perception and distortion: Perception: +12.113; Distortion: +10.358; Total: +11.235 | Hurt |
Surprising findings
- CDC wins on LPIPS-VGG because that metric is part of its training target, while lagging on distortion metrics.
- FID is argued to be unsuitable for compression evaluation since it measures distribution rather than fidelity to the original image.
Limitations
- Relies on a theoretical approximation with several heuristic steps (e.g., Gaussian assumption, gradient-direction approximation).
- Requires running the end-to-end decoder and diffusion sampler at the encoder side, adding encoder complexity.
- Model size of 73.79M parameters across score network, encoder, entropy model and decoder.
Experimental setup
Datasets: DIV2K
Benchmarks: CLIC professional Kodak DIV2K-test
Metrics: PSNR VIF GMSD LPIPS DISTS PIEAPP FID BD rate
Baselines: BPG ELIC HiFiC ILLM CDC
Builds on: DDIM ADM ELIC DISTS LPIPS
Closest papers by meaning
- Decouple Distortion from Perception: Region Adaptive Diffusion for Extreme-low Bitrate Perception Image Compression (CVPR 2025 · similarity 0.90)
- Controllable Distortion-Perception Tradeoff Through Latent Diffusion for Neural Image Compression (AAAI 2025 · similarity 0.90)
- PICD: Versatile Perceptual Image Compression with Diffusion Rendering (CVPR 2025 · similarity 0.90)
- One-Step Diffusion-Based Image Compression with Semantic Distillation (NeurIPS 2025 · similarity 0.88)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.88)
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.