Research map / Image generation / Autoregressive and latent generation
Generalizing Dataset Distillation via Deep Generative Prior
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 |
|---|---|---|---|
| Distill in intermediate generator feature space instead of pixel space | MTT · ImageNet subsets (128×128) | cross-architecture validation accuracy: 33.4±1.5 -> 39.9±1.2 (ImageNet-A); 34.0±3.4 -> 39.4±1.3 (ImageNet-B) | Helped |
| Distill in intermediate generator feature space instead of pixel space | DC · ImageNet subsets (128×128) | cross-architecture validation accuracy: 38.7±4.2 -> 41.8±1.7 (ImageNet-A); 38.7±1.0 -> 42.1±1.2 (ImageNet-B) | Helped |
| Distill in intermediate generator feature space instead of pixel space | DM · ImageNet subsets (128×128) | cross-architecture validation accuracy: 27.2±1.2 -> 31.6±1.4 (ImageNet-A); 24.4±1.1 -> 31.3±3.9 (ImageNet-B) | Helped |
| Distill in intermediate generator feature space instead of pixel space | MTT · CIFAR-10 | average cross-architecture validation accuracy: 24.1±0.8 -> 28.0±0.6 | Helped |
| Distill in intermediate generator feature space instead of pixel space | DC · CIFAR-10 | average cross-architecture validation accuracy: 26.0±0.4 -> 26.3±0.5 (trained generator); 26.0±0.4 -> 26.6±0.5 (random generator) | Mixed |
| Distill in intermediate generator feature space instead of pixel space | DM · CIFAR-10 | average cross-architecture validation accuracy: 22.6±0.5 -> 23.8±0.5 | Helped |
| Use an intermediate F-space rather than an earlier or later latent space (ablation) | MTT + GLaD · ImageNet subsets | validation accuracy | Mixed |
| Use earlier-layer latents rather than later-layer latents (ablation) | MTT + GLaD · ImageNet subsets | cross-architecture validation accuracy | Hurt |
| Use later-layer latents rather than earlier-layer latents (ablation) | MTT + GLaD · ImageNet subsets | cross-architecture validation accuracy | Mixed |
| Initialize latents with Gaussian noise instead of feed-forward initialization (ablation) | GLaD with DC · ImageNet subsets (256×256) | cross-architecture validation accuracy: 32.8±4.1 -> 40.3±1.7 (ImageNet-B, Random G) | Helped |
| Use a randomly initialized generator instead of a dataset-trained generator (ablation) | GLaD with DC · ImageNet subsets (256×256) | cross-architecture validation accuracy: 38.3±4.7 -> 39.3±2.0 (ImageNet-A); 41.5±1.2 -> 40.3±1.7 (ImageNet-B) | Mixed |
| Use generators trained on FFHQ or Pokémon instead of ImageNet (ablation) | GLaD with DC · ImageNet subsets (256×256) | cross-architecture validation accuracy: ImageNet-A: 38.3±4.7 -> 39.1±2.0 (Pokémon); ImageNet-B: 32.8±4.1 -> 40.2±1.1 (FFHQ) | Mixed |
Closest papers by meaning
- Data Distillation Can Be Like Vodka: Distilling More Times For Better Quality (ICLR 2024 · similarity 0.92)
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.91)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.90)
- D^4: Dataset Distillation via Disentangled Diffusion Model (CVPR 2024 · similarity 0.90)
- Large Scale Dataset Distillation with Domain Shift (ICML 2024 · similarity 0.90)
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.