Research map / Image generation / Image restoration
Bridging the Gap between Gaussian Diffusion Models and Universal Quantization for Image Compression
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 quantization-based forward diffusion process using universal quantization, a noise-matched quantization schedule, and a uniform-noise-finetuned latent Stable Diffusion to build a low-bitrate generative image codec.
Problem
Diffusion-based dequantization codecs for image compression suffer from a noise type gap, noise level gap, and discretization gap, so quantized latents fall outside the distribution the diffusion model was trained on, hurting reconstruction quality at very low bitrates.
Approach
Use universal (dithered) quantization in the forward diffusion process with a quantization bin schedule Δt matched to the diffusion variance schedule so SNR matches, and finetune a Gaussian latent diffusion model to uniform noise so the diffusion model matches quantization error distribution.
Key results
- Achieves superior rate-realism performance on MS-COCO 30k below 0.1 bpp and best diffusion-based results on CLIC2020 with performance similar to MS-ILLM.
- Achieves the best LPIPS and MS-SSIM performance below 0.08 bpp on the evaluated datasets.
- Ablations on Kodak show removing uniform-noise diffusion or universal quantization degrades rate-distortion, especially at low bitrates.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use uniform-noise diffusion instead of Gaussian-noise diffusion (ablation) | Proposed codec · Kodak | reconstruction quality | Helped |
| Use universal quantization instead of hard quantization (ablation) | Proposed codec · Kodak | reconstruction quality | Helped |
| Use the proposed codec instead of previous diffusion-based methods | Diffusion-based image compression · MS-COCO 30k below 0.1bpp | rate-realism performance | Helped |
| Use the proposed codec instead of previous diffusion-based methods | Diffusion-based image compression · CLIC2020 | rate-realism performance | Helped |
| Use the proposed codec instead of other evaluated methods | Image compression · Kodak below 0.08bpp | LPIPS | Helped |
Surprising findings
- Uniform-noise diffusion models can be obtained efficiently by finetuning a Gaussian latent diffusion model, despite prior reports that such models could only be trained on toy data.
Limitations
- Possible misgeneration of structural details at low bitrates
- Pixelwise MS-SSIM performance drops above 0.08 bpp relative to baselines
- Uniform diffusion training from scratch was not feasible at practical resolution, so relies on finetuning a Gaussian model
Experimental setup
Datasets: LAION Improved Aesthetics 6.5+ Vimeo90k
Benchmarks: Kodak CLIC2020 MS-COCO 30k
Metrics: FID LPIPS MS-SSIM
Baselines: Relic et al. HFD CDC MR MS-ILLM HiFiC VTM 19.2
Builds on: Stable Diffusion Relic et al. dequantization diffusion codec CompressAI mean-scale hyperprior DDIM universal quantization
Closest papers by meaning
- Progressive Compression with Universally Quantized Diffusion Models (ICLR 2025 · similarity 0.94)
- BitsFusion: 1.99 bits Weight Quantization of Diffusion Model (NeurIPS 2024 · similarity 0.91)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.90)
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.90)
- Optimizing Quantized Diffusion Models via Distillation with Cross-Timestep Error Correction (AAAI 2025 · similarity 0.90)
- Compressed Image Generation with Denoising Diffusion Codebook Models (ICML 2025 · similarity 0.90)
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.