Research map / Image generation / Fast diffusion sampling
Steering One-Step Diffusion Model with Fidelity-Rich Decoder for Fast Image Compression
Zheng Chen, Mingde Zhou, Jinpei Guo, Jiale Yuan, Yifei Ji, Yulun Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add fidelity guidance module (aux. fidelity guidance) (ablation) | SODEC · DIV2K-Val | MS-SSIM: 0.8212 -> 0.8481 | Improved |
| add fidelity guidance module (ablation) | SODEC · DIV2K-Val | LPIPS: 0.3625 -> 0.3351 | Improved |
| use text prompt guidance instead of none (ablation) | SODEC · DIV2K-Val | MS-SSIM: 0.8212 -> 0.8185 | Hurt |
| use hyperprior guidance instead of none (ablation) | SODEC · DIV2K-Val | MS-SSIM: 0.8212 -> 0.8258 | Improved |
| add MSE alignment loss (ablation) | SODEC · DIV2K-Val | LPIPS: 0.4210 -> 0.3827 | Improved |
| use MSE+LPIPS composite alignment loss instead of MSE only (ablation) | SODEC · DIV2K-Val | MS-SSIM: 0.7948 -> 0.7481 | Hurt |
| merge alignment into main loss instead of separate term (ablation) | SODEC · DIV2K-Val | LPIPS: 0.3827 -> 0.4023 | Hurt |
| use rate annealing instead of frozen VAE (ablation) | SODEC · DIV2K-Val | MS-SSIM: 0.8512 -> 0.8951 | Improved |
| use rate annealing instead of joint training at matched bpp (ablation) | SODEC · DIV2K-Val | LPIPS: 0.3750 -> 0.3113 | Improved |
| use high-to-low rate annealing instead of low-to-high bpp curriculum (ablation) | SODEC · DIV2K-Val | MS-SSIM: 0.8643 -> 0.8951 | Improved |
| one-step diffusion decoding instead of multi-step PerCo | SODEC · DIV2K-Val | decoding time (ms): 4,702.2 -> 227.9 | Improved |
| one-step diffusion vs VAE-only HiFiC decoding | SODEC · DIV2K-Val | decoding time (ms): 3.9 -> 227.9 | Hurt |
Closest papers by meaning
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.91)
- Lossy Compression with Pretrained Diffusion Models (ICLR 2025 · similarity 0.90)
- Correcting Diffusion-Based Perceptual Image Compression with Privileged End-to-End Decoder (ICML 2024 · similarity 0.90)
- One-Step Diffusion-Based Image Compression with Semantic Distillation (NeurIPS 2025 · similarity 0.90)
- Decouple Distortion from Perception: Region Adaptive Diffusion for Extreme-low Bitrate Perception Image Compression (CVPR 2025 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · 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.
