Research map / Label-efficient and robust learning / Recognition and long-tail
Demystifying Randomly Initialized Networks for Evaluating Generative Models
Junghyuk Lee, Jun-Hyuk Kim, Jong-Seok Lee
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 random Swin features rather than trained SwAV features for FID evaluation | Generative models across ImageNet, FFHQ, and Pokémon · ImageNet 256², ImageNet 64², FFHQ 1024²/256², Pokémon 256² | FID model ranking | Mixed |
| Use random features to assess low-level image differences such as color, blur, and noise | FID evaluation of disturbed ImageNet images · ImageNet validation | FID sensitivity to low-level disturbance: Gaussian noise level 3: Inception (trained) 41.09; Swin (random) 0.16; color jitter level 3: Inception (trained) 28.16; | Improved |
| Use trained features to assess high-level semantic changes rather than random features alone | FID evaluation of outlier-replaced FFHQ sets · FFHQ | FID sensitivity to high-level outliers | Mixed |
| Combine trained and random features to capture distinct types of generated-image abnormalities | Generative-model evaluation · FFHQ and ImageNet | coverage of high- and low-level outliers: 0.276, 0.238, 0.152; 0.191 and 0.127 | Improved |
| Use Transformer rather than CNN random features | Random-feature FID evaluation · Generative models across datasets | variation across random seeds | Improved |
| Use Swin-T or ViT-T instead of larger tested models (ablation) | Random-feature FID evaluation · ImageNet 256² | model ranking consistency: BigGAN: Swin-T 1.07, Swin-S 1.07, Swin-B 1.54; ViT-T 0.62, ViT-B 3.14 | No clear effect |
| Increase Swin size from T to S (ablation) | Random-feature evaluation of BigGAN and ADM-U · ImageNet 256² | FID scale: BigGAN: 1.07 -> 1.07; ADM-U: 3.57 -> 3.70 | No clear effect |
| Increase Swin size from T to B (ablation) | Random-feature evaluation of BigGAN and ADM-U · ImageNet 256² | FID scale: BigGAN: 1.07 -> 1.54; ADM-U: 3.57 -> 5.03 | Mixed |
| Increase ViT size from T to B (ablation) | Random-feature evaluation of BigGAN and ADM-U · ImageNet 256² | FID scale: BigGAN: 0.62 -> 3.14; ADM-U: 1.61 -> 6.61 | Mixed |
| Average random-feature results across different initialization seeds | Random-feature FID evaluation · Generative models across datasets | variation across seeds | Improved |
Closest papers by meaning
- Generative Modeling of Weights: Generalization or Memorization? (CVPR 2026 · similarity 0.87)
- A Theoretical Framework for Statistical Evaluability of Generative Models (ICML 2026 · similarity 0.86)
- On the Power and Limitations of Random Features for Understanding Neural Networks (NeurIPS 2019 · similarity 0.86)
- Assessing Generative Models via Precision and Recall (NeurIPS 2018 · similarity 0.86)
- Biased Generalization in Diffusion Models (ICML 2026 · similarity 0.85)
- Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space (CVPR 2017 · 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.
