Research map / Image generation / Autoregressive and latent generation
Implicit Neural Representation Facilitates Unified Universal Vision Encoding
Matthew Gwilliam, Xiao Wang, Xuefeng Hu, Zhenheng Yang
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 |
|---|---|---|---|
| make INR hyper-network patch-wise (ablation) | TransINR hyper-network · ImageNette reconstruction | PSNR: 27.15 -> 51.96 | Improved |
| add global token (ablation) | HUVR hyper-network · ImageNette reconstruction | PSNR: 51.96 -> 53.36 | Improved |
| add TinTok compression (ablation) | HUVR hyper-network · ImageNette reconstruction | PSNR: 53.36 -> 48.58 | Hurt |
| distill from DINOv3 to global token only vs none (ablation) | HUVR ViT-B · ImageNet linear probe d=768 | accuracy: 11.4 -> 82.9 | Improved |
| distill to patch tokens (ablation) | HUVR ViT-B · ADE20K | mIoU: 4.18 -> 45.23 | Improved |
| distill to both global and patch tokens (ablation) | HUVR ViT-B · ImageNet reconstruction | PSNR: 28.42 -> 25.33 | Hurt |
| add INR reconstruction objective (ablation) | HUVR ViT-B · ImageNet TinToks d=32 | accuracy: 64.2 -> 68.2 | Improved |
| remove attention from TinTok decoder (ablation) | HUVR ViT-B · TinToks classification vs reconstruction | accuracy: 69.7 -> 73.5 acc; PSNR 25.35 -> 24.90 | Mixed |
| double INR hidden dim to 512 (ablation) | HUVR ViT-B · ADE20K | mIoU: 43.85 -> 23.36 | Hurt |
| use larger distillation teacher (ViT-L vs ViT-B) (ablation) | HUVR ViT-B · ImageNet linear probe | accuracy: 82.5 -> 80.3 at 10 ep; 84.1 -> 84.6 at 50 ep | Mixed |
| use HUVR instead of DINOv3 for ViT-B | ViT-B/16 · ImageNet linear probe | accuracy: 84.6 -> 85.0 | Improved |
| use HUVR TinToks instead of DINOv3 PCA at 32 dims | ViT-B · ImageNet-1k linear probe | accuracy: 64.1 -> 79.4 | Improved |
Closest papers by meaning
- Polynomial Implicit Neural Representations for Large Diverse Datasets (CVPR 2023 · similarity 0.88)
- Implicit Representations via Operator Learning (ICML 2024 · similarity 0.86)
- Towards flexible perception with visual memory (ICML 2025 · similarity 0.86)
- HIIF: Hierarchical Encoding based Implicit Image Function for Continuous Super-resolution (CVPR 2025 · similarity 0.85)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.85)
- Explaining the Implicit Neural Canvas: Connecting Pixels to Neurons by Tracing their Contributions (CVPR 2024 · 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.
