Research map / Image generation / Image restoration
Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis
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 a neural image codec with a non-isotropic blurring-diffusion decoder conditioned on a quantized latent, plus a channel-wise entropy model using a Transformer with learnable Laplacian-shaped relative positional encoding per channel chunk.
Problem
Diffusion-based decoders in learned image compression lack image-specific inductive bias, limiting perceptual quality; entropy models also trade off compression efficiency against decoding speed.
Approach
Replace the isotropic DDPM decoder with a blurring diffusion model that diffuses frequency components at different rates (coarse-to-fine generation), and pair it with an uneven channel-chunked entropy model whose Transformer global context uses a learnable Laplacian relative position encoding adapted per chunk.
Key results
- Achieves better rate-perception (FID, LPIPS) than HiFiC, Multi-Realism, DIRAC and CDC on Kodak and CLIC2020, at the cost of PSNR relative to most baselines (except CDC).
- Laplacian-shaped positional encoding gives 0.2347 bpp vs 0.2512 for diamond relative position encoding and 0.2811 with no positional encoding in the ablation at lambda=0.01.
- The parallel local+global entropy model achieves -8.25% BD-rate vs VVC on Kodak with 196.3 ms decoding latency, versus -7.89% for MEM at 207.4 ms.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase maximum blur from σB,max=0 to σB,max=25 (ablation) | Blurring diffusion decoder · Kodak | FID: 3.94 -> 3.45 | Helped |
| set maximum blur to σB,max=5 instead of σB,max=0 (ablation) | Blurring diffusion decoder · Kodak | FID: 3.94 -> 3.78 | Helped |
| set maximum blur to σB,max=15 instead of σB,max=0 (ablation) | Blurring diffusion decoder · Kodak | FID: 3.94 -> 3.56 | Helped |
| add relative positional encoding (ablation) | Transformer entropy model · Kodak | bpp: 0.2811 -> 0.2643 | Helped |
| replace relative positional encoding with diamond relative positional encoding (ablation) | Transformer entropy model · Kodak | bpp: 0.2643 -> 0.2512 | Helped |
| add Laplacian-shaped positional encoding instead of diamond relative positional encoding (ablation) | Transformer entropy model · Kodak | bpp: 0.2512 -> 0.2347 | Helped |
| use parallel instead of serial local and global context (ablation) | Proposed entropy model · Kodak | Decoding latency (ms): > 103 -> 196.3 | Helped |
| use parallel instead of serial local and global context (ablation) | Proposed entropy model · Kodak | BD-Rate: -9.38 -> -8.25 | Hurt |
| add global spatial context to local spatial context | Proposed entropy model · Kodak | BD-Rate: -5.71 -> -7.89 | Helped |
Surprising findings
- Their model is the only one besides CDC with lower PSNR; it trades distortion for realism.
Limitations
- Distortion (PSNR) is worse than most baselines; the paper accepts this for realism.
- Serial global-context variant exceeds 103 ms decoding; only the parallel variant is fast.
Experimental setup
Datasets: DIV2K Flickr2K CLIC
Benchmarks: Kodak CLIC2020
Metrics: PSNR FID LPIPS BD-rate bpp inference latency
Baselines: HiFiC Multi-Realism DIRAC CDC VVC-Intra (VTM) DGML NSC MEM
Builds on: blurring diffusion model CDC Entroformer ELIC Swin Transformer
Closest papers by meaning
- Decouple Distortion from Perception: Region Adaptive Diffusion for Extreme-low Bitrate Perception Image Compression (CVPR 2025 · similarity 0.89)
- Diversify, Contextualize, and Adapt: Efficient Entropy Modeling for Neural Image Codec (NeurIPS 2024 · similarity 0.88)
- Controllable Distortion-Perception Tradeoff Through Latent Diffusion for Neural Image Compression (AAAI 2025 · similarity 0.88)
- Correcting Diffusion-Based Perceptual Image Compression with Privileged End-to-End Decoder (ICML 2024 · similarity 0.87)
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.87)
- Compressed Image Generation with Denoising Diffusion Codebook Models (ICML 2025 · similarity 0.87)
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.