Research map / Image generation / Autoregressive and latent generation
Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models
George Stein, Jesse C. Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecroze, Zhaoyan Liu, Anthony L. Caterini, Eric Taylor, Gabriel Loaiza-Ganem
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 |
|---|---|---|---|
| replace Inception-V3 encoder with DINOv2-ViT-L/14 in Frechet distance (ablation) | FD-based metric · CIFAR10, ImageNet, FFHQ, LSUN-Bedroom | correlation with human error rate | Improved |
| use FID (Inception) to rank models | diffusion models vs GANs · ImageNet, LSUN, FFHQ | agreement with human realism | Hurt |
| use SwAV or CLIP-B/32 instead of Inception as FD encoder (ablation) | FD metric · human alignment | correlation with human error rate | No clear effect |
| use CLIP ViT-L/14, DINOv2 or MAE instead of Inception as FD encoder (ablation) | FD metric · human alignment | correlation with human error rate | Improved |
| use precision instead of FD as fidelity measure (ablation) | fidelity metric · human alignment | correlation with human error rate | Hurt |
| measure diversity with per-class Vendi score to test whether low diversity explains FID mismatch (ablation) | ImageNet models · ImageNet | explanation of FID vs human mismatch | No clear effect |
| use memorization metrics AuthPct, C_T and FLS-POG to detect memorization (ablation) | memorization metrics · CIFAR10 | tracking memorization ratio | Hurt |
Closest papers by meaning
- A Unifying Information-theoretic Perspective on Evaluating Generative Models (AAAI 2025 · similarity 0.89)
- Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion (ICML 2025 · similarity 0.89)
- FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance (NeurIPS 2026 · similarity 0.88)
- Attribute Based Interpretable Evaluation Metrics for Generative Models (ICML 2024 · similarity 0.88)
- FlashEval: Towards Fast and Accurate Evaluation of Text-to-image Diffusion Generative Models (CVPR 2024 · similarity 0.88)
- Revisiting text-to-image evaluation with Gecko: on metrics, prompts, and human rating (ICLR 2025 · similarity 0.88)
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.
