Research map / Image generation / Fast diffusion sampling
One-Step Diffusion-Based Image Compression with Semantic Distillation
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 OneDC, a one-step diffusion image codec that uses the compressed hyperprior (with semantic distillation from a generative tokenizer) as cross-attention guidance instead of text, trained in two stages with hybrid pixel/latent losses.
Problem
Multi-step diffusion-based generative image codecs achieve strong perceptual quality at low bitrates but suffer high decoding latency from iterative sampling, and text prompts are poor semantic guidance for compression.
Approach
Replace text prompts with the categorical (FSQ-quantized) hyperprior as semantic conditioning for a one-step diffusion generator, enhance it via distillation of token labels from a pretrained MaskGIT tokenizer, and train with hybrid pixel-domain (L1+LPIPS) and latent-domain diffusion distillation with adversarial loss.
Key results
- OneDC achieves 55.27% LPIPS bitrate saving over DiffC on Kodak and 54.60% on CLIC2020.
- OneDC gives 39.55% bitrate saving over PerCo (SD) on MS-COCO 30K.
- OneDC decodes over 20x faster than multi-step diffusion codecs (0.34 s vs 6.9-12.4 s on 1024x1024, A100).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove semantic guidance (ablation) | OneDC · CLIC2020 test set | DISTS BD-Rate (%) ↓: 0.00 -> 44.0 | Hurt |
| replace text guidance with hyperprior guidance (ablation) | OneDC · CLIC2020 test set | DISTS BD-Rate (%) ↓: 24.2 -> 20.7 | Helped |
| add semantic distillation to hyperprior guidance (ablation) | OneDC · CLIC2020 test set | DISTS BD-Rate (%) ↓: 20.7 -> 0.00 | Helped |
| replace text guidance with hyperprior guidance (ablation) | OneDC · CLIC2020 test set | FID BD-Rate (%) ↓: 36.3 -> 24.3 | Helped |
| add semantic distillation to hyperprior guidance (ablation) | OneDC · CLIC2020 test set | FID BD-Rate (%) ↓: 24.3 -> 0.00 | Helped |
| add latent-domain distillation and adversarial losses (ablation) | OneDC pixel-domain training · CLIC2020 test set | FID BD-Rate (%) ↓: 51.8 -> 0.00 | Helped |
| add pixel-domain reconstruction loss (ablation) | OneDC latent-domain training · CLIC2020 test set | DISTS BD-Rate (%) ↓: 60.7 -> 0.00 | Helped |
| increase latent bitrate from 0.0 to 0.0157 bpp while holding hyperprior at 0.0035 bpp (ablation) | OneDC · CLIC2020 test set | LPIPS: 0.290 -> 0.119 | Helped |
| use OneDC instead of DiffEIC | image decoding on MS-COCO 30K · MS-COCO 30K, 1024×1024 images | decoding time (s): 12.4 -> 0.34 | Helped |
| use OneDC instead of PerCo (SD) | image compression · MS-COCO 30K | FID BD-Rate (%) ↓: 59.6 -> 0.00 | Helped |
Surprising findings
- Text-prompt guidance underperforms a compressed hyperprior as conditioning for one-step diffusion decoding.
- With zero bits for the latent, the hyperprior alone (0.0035 bpp) still yields coherent reconstructions.
Limitations
- Decoding speed does not meet real-time requirements.
- Higher memory usage (8,038 MB) than DDCM on Kodak.
- Large model size (1.4B parameters total).
Experimental setup
Datasets: training dataset from prior work (unspecified)
Benchmarks: Kodak CLIC2020 MS-COCO 30K DIV2K
Metrics: LPIPS DISTS PSNR MS-SSIM FID BD-Rate
Baselines: MS-ILLM GLC PerCo (SD) DiffEIC DiffC DDCM
Builds on: DiffEIC GLC DMD MaskGIT LoRA
Closest papers by meaning
- Bridging the Gap between Gaussian Diffusion Models and Universal Quantization for Image Compression (CVPR 2025 · similarity 0.90)
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.89)
- PICD: Versatile Perceptual Image Compression with Diffusion Rendering (CVPR 2025 · similarity 0.89)
- Ultra Lowrate Image Compression with Semantic Residual Coding and Compression-aware Diffusion (ICML 2025 · similarity 0.89)
- DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence (NeurIPS 2026 · similarity 0.89)
- One-Way Ticket: Time-Independent Unified Encoder for Distilling Text-to-Image Diffusion Models (CVPR 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.