Research map / Image generation / Fast diffusion sampling
OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates
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 OSCAR, a one-step diffusion image codec that maps compression bit-rates to pseudo diffusion timesteps so one Stable Diffusion 2.1-based model with LoRA and bit-rate-specific VQ hyper-encoders reconstructs images at multiple bit-rates in a single denoising pass.
Problem
Diffusion-based image codecs need many iterative denoising steps and separate models per bit-rate, making them computationally expensive and storage-heavy.
Approach
Treat quantized latents as noisy intermediate diffusion states whose noise level is set by bit-rate; align quantized latents to original latents via cosine similarity, map each bit-rate to a pseudo diffusion timestep through the measured similarity matching sqrt(alpha_bar_t), and fine-tune a single denoiser to reconstruct in one step.
Key results
- OSCAR consistently outperforms listed baselines on LPIPS, DISTS, MUSIQ and CLIPIQA across Kodak, DIV2K-Val and CLIC2020 according to the paper's comparison figures.
- OSCAR reaches competitive performance with multi-step diffusion codecs while running orders of magnitude faster in a single denoising step.
- A new unseen bit-rate can be adapted with 15k extra iterations and only minor degradation on existing rates.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use calibrated bit-rate-to-timestep mapping instead of random timestep assignment (ablation) | OSCAR · Kodak, across bit-rates | evaluation metrics | Helped |
| add representation alignment (RePA) (ablation) | OSCAR · Kodak | evaluation metrics | Helped |
| add edge-aware DISTS loss (ablation) | OSCAR · Kodak | evaluation metrics | Helped |
| add GAN loss (ablation) | OSCAR · Kodak | evaluation metrics | Helped |
Surprising findings
- Diffusion-based codecs, including OSCAR, underperform traditional codecs in PSNR, partly because the VAE alone at high bit-rates yields lower PSNR than BPG/VVC.
Limitations
- RD-curves are not strictly convex because bit-rate depends on both downsampling ratio and codebook size.
- PSNR lags behind traditional codecs due to VAE reconstruction limits.
- Multiple hyper-encoders are still trained per predefined bit-rate, so only a fixed set of rates is directly supported.
Experimental setup
Datasets: DIV2K Flickr2K LSDIR Kodak DIV2K-val CLIC2020
Benchmarks: Kodak DIV2K-val CLIC2020
Metrics: LPIPS DISTS MUSIQ CLIPIQA PSNR FID BD-rate
Baselines: BPG VVC HiFiC MS-ILLM DPICT CTC CDC DiffEIC PerCo
Builds on: Stable Diffusion 2.1 VQ-VAE DISTS LoRA
Closest papers by meaning
- Progressive Compression with Universally Quantized Diffusion Models (ICLR 2025 · similarity 0.91)
- Compressed Image Generation with Denoising Diffusion Codebook Models (ICML 2025 · similarity 0.91)
- Bridging the Gap between Gaussian Diffusion Models and Universal Quantization for Image Compression (CVPR 2025 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.90)
- Lossy Compression with Pretrained Diffusion Models (ICLR 2025 · similarity 0.90)
- One-Step Diffusion-Based Image Compression with Semantic Distillation (NeurIPS 2025 · similarity 0.89)
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.