Research map / Vision-language and multimodal / CLIP and zero-shot
NAF: Zero-Shot Feature Upsampling via Neighborhood Attention Filtering
Loïck Chambon, Paul Couairon, Éloi Zablocki, Alexandre Boulch, Nicolas Thome, Matthieu Cord
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 |
|---|---|---|---|
| use NAF instead of VFM-specific JAFAR | upsampler · Pascal VOC12 | mIoU gain over Nearest: +5.12 -> +5.58 | Improved |
| add context (3x3) branch to the pixel encoder (ablation) | NAF encoder · Kitti360 | mIoU: 59.86 -> 60.41 | Improved |
| use dual-branch blocks instead of Inception-style blocks (ablation) | NAF encoder · Kitti360 | mIoU: 59.56 -> 60.41 | Improved |
| increase guidance dimension C from 256 to 1024 (ablation) | NAF encoder · Kitti360 | mIoU: 60.41 -> 61.05 | Improved |
| increase conv blocks L from 2 to 5 (ablation) | NAF encoder · Kitti360 | mIoU: 60.41 -> 61.13 | Improved |
| add 1x1 conv after pooling of keys (ablation) | NAF attention keys · Cityscapes | mIoU: 60.41 -> 58.38 | Hurt |
| pool keys (AvgPool) instead of bilinear interpolation (ablation) | NAF attention keys · Cityscapes | mIoU: 59.66 -> 60.41 | Improved |
| add RoPE positional encoding (ablation) | NAF attention · Cityscapes | mIoU: 57.07 -> 60.41 | Improved |
| use RoPE instead of Gaussian spatial kernel (ablation) | NAF attention · Cityscapes | mIoU: 58.19 -> 60.41 | Improved |
| use NAF instead of JAFAR for video propagation | upsampler · DAVIS | J&F gain: +3.54 vs +3.37 | Mixed |
| replace bilinear with NAF in ProxyCLIP | open-vocab segmentation · Pascal VOC | mIoU: +1.04 mean gain | Improved |
| use NAF for Gaussian denoising instead of Restormer | image restoration · ImageNet Gaussian denoising sigma=0.1 | PSNR: 32.51 vs 32.12 | 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.
- JAFAR — JAFAR: Jack up Any Feature at Any Resolution (NeurIPS 2025) · better on 6 of 7 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
- LoftUp: Learning a Coordinate-Based Feature Upsampler for Vision Foundation Models (ICCV 2025 · similarity 0.88)
- JAFAR: Jack up Any Feature at Any Resolution (NeurIPS 2025 · similarity 0.86)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.84)
- TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning (NeurIPS 2024 · similarity 0.84)
- LR0.FM: Low-Res Benchmark and Improving robustness for Zero-Shot Classification in Foundation Models (ICLR 2025 · similarity 0.84)
- Zero-to-Hero: Enhancing Zero-Shot Novel View Synthesis via Attention Map Filtering (NeurIPS 2024 · similarity 0.84)
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.
