Research map / Image generation / Autoregressive and latent generation
FunkNN: Neural Interpolation for Functional Generation
AmirEhsan Khorashadizadeh, Anadi Chaman, Valentin Debarnot, Ivan Dokmanić
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 |
|---|---|---|---|
| Pruning LIIF-EDSR from 1.5M to 140K parameters | LIIF-EDSR · CelebA-HQ 128→256 and 128→512 | SNR (dB): 31.6→31.6 at 256; 28.0→27.5 at 512 | Hurt |
| Patch-based FunkNN with factor training (140K params) vs. LIIF-EDSR (1.5M params) | FunkNN · CelebA-HQ 128→256/512/1024 | SNR (dB): 32.6 vs 31.6 at 256; 28.2 vs 28.0 at 512; 24.2 vs 23.9 at 1024 | Mixed |
| Factor training strategy instead of single-scale training (ablation) | FunkNN · CelebA-HQ in-distribution 128→256/512/1024 | SNR (dB): +1.4 at 256, +1.9 at 512, +1.7 at 1024 over single | Improved |
| Factor training strategy instead of single-scale training (ablation) | FunkNN · LSUN-bedroom OOD 128→256 | SNR (dB): 26.6 vs 27.2 (-0.6) | Hurt |
| Continuous training strategy (multiple input scales) instead of single (ablation) | FunkNN · CelebA-HQ 128→512 and 128→1024 | SNR (dB): +1.4 at 512, +1.0 at 1024 | Improved |
| Continuous training strategy instead of single (ablation) | FunkNN · CelebA-HQ 128→256 and LSUN OOD 128→256 | SNR (dB): -0.5 at 256; -0.8 on OOD | Hurt |
| Patch-based FunkNN (factor) vs U-Net trained for 128→256 | Super-resolution to OOD LSUN-bedroom · LSUN-bedroom 128→256 (OOD) | SNR (dB): 26.6 vs 19.9 (+6.7) | Improved |
| U-Net super-resolution baseline vs FunkNN on inlier data | CelebA-HQ 128→256 | SNR (dB): 31.3 vs 32.6 for FunkNN (factor) | Hurt |
| Patch-based FunkNN generalization to OOD data (LSUN-bedroom) vs LIIF-EDSR at 1.5M params | FunkNN · LSUN-bedroom OOD 128→256 | SNR (dB): 26.6 vs 28.0 | Hurt |
| Inference speed: FunkNN vs LIIF-EDSR (1.5M) | FunkNN · 100 images of 128x128 on A100 | time (s): 0.7 s vs 1.97 s | Improved |
Closest papers by meaning
- DEFT: Efficient Fine-tuning of Diffusion Models by Learning the Generalised $h$-transform (NeurIPS 2024 · similarity 0.85)
- pcaGAN: Improving Posterior-Sampling cGANs via Principal Component Regularization (NeurIPS 2024 · similarity 0.85)
- Consistency Posterior Sampling for Diverse Image Synthesis (CVPR 2025 · similarity 0.85)
- Stochastic Interpolants via Conditional Dependent Coupling (NeurIPS 2026 · similarity 0.85)
- Generalization in diffusion models arises from geometry-adaptive harmonic representations (ICLR 2024 · similarity 0.85)
- Solving Inverse Problems with FLAIR (NeurIPS 2025 · 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.
