Research map / Image generation / Autoregressive and latent generation
TiltDiff: Tilted Weight-Space Diffusion for Neural Network Generation
En-Ni Chuang, Hanjuan Huang, Hao-Jia Song, Hsing-Kuo Kenneth Pao, Tyng-Luh Liu
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 |
|---|---|---|---|
| predict injected noise instead of clean latent z0 (ablation) | TiltDiff · CIFAR-10 zoo | Top-3 accuracy: 21.17 -> 42.18 | Improved |
| add validation-accuracy weighting to noise-prediction diffusion (ablation) | TiltDiff · CIFAR-10 zoo | Top-3 accuracy: 42.18 -> 46.76 | Improved |
| add validation-based data selection (ablation) | TiltDiff · CIFAR-10 zoo | Top-3 accuracy: 42.18 -> 58.70 | Improved |
| add accuracy weighting on top of data selection (ablation) | TiltDiff · CIFAR-10 zoo | Top-3 accuracy: 58.70 -> 60.76 | Improved |
| add sigmoid accuracy weighting vs SANESub | TiltDiffSig · CIFAR-10, top 10% retained | Top-3 accuracy: 56.79 -> 60.76 | Improved |
| replace SANE with diffusion on top-30% subset | TiltDiff · CIFAR-10, top 30% retained | Top-3 accuracy: 57.37 -> 58.14 | Improved |
| accuracy weighting on CIFAR-100 heads at 35% retained (exp weighting) (ablation) | TiltDiffExp · CIFAR-100 ResNet-18 heads, 35% retained | Top-3 accuracy: 70.94 -> 69.09 | Hurt |
| accuracy weighting on CIFAR-10, 25% retained (sigmoid) (ablation) | TiltDiffSig · CIFAR-10, top 25% retained | Top-3 accuracy: 56.58 -> 55.69 | Hurt |
| use diffusion-generated weights vs SANESub as initialization | TiltDiff · CIFAR-10 fine-tuning, 0 epochs | test accuracy: 57.9 -> 60.8 | Improved |
| use TiltDiff init vs D2NWG after 25 epochs on MNIST | TiltDiff · MNIST fine-tuning, 25 epochs | test accuracy: 96.2 -> 91.3 | Hurt |
| use TiltDiff init vs D2NWG after 25 epochs on CIFAR-10 | TiltDiff · CIFAR-10 fine-tuning, 25 epochs | test accuracy: 64.1 -> 69.6 | Improved |
| mask Block 4 cat pathway | TiltDiff sampled model · CIFAR-10 cat | cat accuracy: 0.80 (-44.00) | Hurt |
Closest papers by meaning
- Diffusion-based Neural Network Weights Generation (ICLR 2025 · similarity 0.88)
- Improving equilibrium propagation without weight symmetry through Jacobian homeostasis (ICLR 2024 · similarity 0.86)
- Towards Scalable and Versatile Weight Space Learning (ICML 2024 · similarity 0.86)
- Weight Diffusion for Future: Learn to Generalize in Non-Stationary Environments (NeurIPS 2024 · similarity 0.85)
- Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models (NeurIPS 2025 · similarity 0.85)
- From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks (ICLR 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.
