Research map / Image generation / Image restoration
Ultra-Low Bitrate Perceptual Image Compression with Shallow Encoder
Tianyu Zhang, Dong Liu, Chang Wen Chen
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 encoder-side distillation (Lenc) when training the shallow encoder (ablation) | AEIC-SE · DIV2K (768x512 patches) | BD-rate DISTS vs AEIC-ME: +23.75 -> +7.68 | Improved |
| add decoder-side distillation (Ldec) on top of Lenc (ablation) | AEIC-SE · DIV2K (768x512 patches) | BD-rate DISTS vs AEIC-ME: +7.68 -> +2.55 | Improved |
| add encoder distillation (ablation) | AEIC-SE · DIV2K (768x512 patches) | BD-rate FID: +22.10 -> +4.75 | Improved |
| use 32x spatial compression instead of 64x (ablation) | AEIC-ME · Kodak | BD-rate DISTS vs StableCodec: -6.95 -> -24.91 | Improved |
| use 32x spatial compression instead of 16x (ablation) | AEIC-ME · Kodak | BD-rate PSNR/LPIPS vs StableCodec: PSNR -5.53 -> -2.21; LPIPS -1.91 -> -13.67 | Mixed |
| use 64x spatial compression instead of 32x (ablation) | AEIC-ME · Kodak | BD-rate PSNR vs StableCodec: -2.21 -> +14.30 | Hurt |
| add high-resolution finetuning stage (5K iters, 1024 patches) (ablation) | AEIC-SE · DIV2K (full 2K resolution) | DISTS | Improved |
| add high-resolution finetuning stage (ablation) | AEIC-SE · DIV2K (768x512 patches) | LPIPS/DISTS | No clear effect |
| replace shallow 0.94M encoder with moderate encoder (AEIC-ME) | AEIC · RTX 4090D 1080P encoding | encoding latency (ms): 27.9 -> 58.7 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- StableCodec — StableCodec: Taming One-Step Diffusion for Extreme Image Compression (ICCV 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Towards image compression with perfect realism at ultra-low bitrates (ICLR 2024 · similarity 0.91)
- Decouple Distortion from Perception: Region Adaptive Diffusion for Extreme-low Bitrate Perception Image Compression (CVPR 2025 · similarity 0.89)
- Ultra Lowrate Image Compression with Semantic Residual Coding and Compression-aware Diffusion (ICML 2025 · similarity 0.88)
- Generative Latent Coding for Ultra-Low Bitrate Image Compression (CVPR 2024 · similarity 0.87)
- Compress Clean Signal from Noisy Raw Image: A Self-Supervised Approach (ICML 2024 · similarity 0.85)
- Towards Unified Human Perception and Machine Understanding: Token Flow Guided Compression Framework (CVPR 2026 · similarity 0.85)
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.
