Research map / Image generation / Image restoration
Improving Statistical Fidelity for Neural Image Compression with Implicit Local Likelihood Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace PatchGAN with the VQ-VAE-conditioned local ILLM discriminator | Mean-Scale Hyperprior neural image compression · CLIC2020 test | FID | Helped |
| Replace PatchGAN with the VQ-VAE-conditioned local ILLM discriminator | Mean-Scale Hyperprior neural image compression · CLIC2020 test | PSNR | No clear effect |
| Add adversarial fine-tuning | No-GAN image compression · CLIC2020 text example | text legibility | Hurt |
| Use no normalization instead of instance normalization in the U-Net discriminator (ablation) | MS-ILLM · CLIC2020 test, bitrate ≈ 0.035 | FID: 8.28 -> 6.27 | Helped |
| Use no normalization instead of instance normalization in the U-Net discriminator (ablation) | MS-ILLM · CLIC2020 test, bitrate ≈ 0.121 | FID: 3.04 -> 2.65 | Helped |
| Use no normalization instead of instance normalization in the U-Net discriminator (ablation) | MS-ILLM · CLIC2020 test, bitrate ≈ 0.231 | FID: 2.21 -> 1.77 | Helped |
| Use no normalization instead of spectral normalization in the U-Net discriminator (ablation) | MS-ILLM · CLIC2020 test | training speed: 25 times | Helped |
| Vary labeler spatial size and codebook size (ablation) | MS-ILLM · CLIC2020 test, bitrate ≈ 0.121 | FID: 2.66, 2.65, 2.57, 2.55 | No clear effect |
| Use linear warmup for 10,000 steps followed by cosine learning-rate decay (ablation) | Mean-Scale Hyperprior pretraining · CLIC2020 test | PSNR | Helped |
| Include LPIPS when training the VQ-VAE labeler (ablation) | VQ-VAE labeler for MS-ILLM · Image compression | labeler success | Helped |
| Increase adversarial loss weight λd (ablation) | MS-ILLM generator fine-tuning · Image compression rate-distortion-perception tradeoff | FID | Helped |
| Increase adversarial loss weight λd (ablation) | MS-ILLM generator fine-tuning · Image compression rate-distortion-perception tradeoff | MSE | Hurt |
Closest papers by meaning
- End-to-end Optimized Image Compression (ICLR 2017 · similarity 0.90)
- Improving Inference for Neural Image Compression (NeurIPS 2020 · similarity 0.89)
- High-Fidelity Generative Image Compression (NeurIPS 2020 · similarity 0.88)
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.88)
- Hierarchical Quantized Autoencoders (NeurIPS 2020 · similarity 0.88)
- Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion (CVPR 2025 · similarity 0.87)
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.