Research map / Probabilistic and scientific modeling / Variational and generative sampling
Unveiling Differences in Generative Models: A Scalable Differential Clustering Approach
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
x
Problem
Comparing two generative models at a fine-grained level requires identifying which sample types one model produces more frequently than a reference; existing kernel spectral methods (KEN) and novelty scores scale poorly to large sample sizes and many modes.
Approach
Approximates the eigenspace of the ρ-conditional kernel covariance difference (C_X − ρC_Y) for a Gaussian kernel using r random Fourier features, computing a 2r×2r matrix via stochastic averaging, so cost depends on r rather than sample size; includes an eigenvalue/eigenvector perturbation bound for this non-symmetric setting.
Key results
- FINC runs on ImageNet-1K with 250k samples in 3090 s on CPU and 74.3 s on GPU, while baselines run out of memory beyond 50k–10k samples.
- FINC with r=2000 Fourier features keeps approximation error below 1% in grid search.
- FINC recovers the ten colored digit types as top-10 novel modes in a colored MNIST sanity check.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FINC (r = 2000) instead of the baseline differential-clustering methods | ImageNet differential clustering · ImageNet | runtime (seconds): FINC (r = 2000): 0.5 to 74.3 (GPU); KEN: 1.1 to 455.6; KEN (Cholesky): 0.1 to 16.7; Rarity: 0.1 to 27.6; FLD: 0.2 to 25. | Mixed |
| Use FINC (r = 2000) for larger sample sizes | ImageNet differential clustering · ImageNet, 25k–250k samples | runtime (seconds): FINC (GPU): 7.6, 15.0, 29.8, 74.3; FINC (CPU): 311.2, 617.6, 1248, 3090 | Helped |
Limitations
- Only pairwise comparison of two models is supported.
- Results are largely qualitative visualizations rather than quantitative benchmarks.
Experimental setup
Baselines: Rarity score FLD score KEN KEN (Cholesky)
Builds on: KEN (kernel-based entropic novelty) random Fourier features
Closest papers by meaning
- An Interpretable Evaluation of Entropy-based Novelty of Generative Models (ICML 2024 · similarity 0.88)
- Towards a Scalable Reference-Free Evaluation of Generative Models (NeurIPS 2024 · similarity 0.87)
- Image Generation Diversity Issues and How to Tame Them (CVPR 2025 · similarity 0.86)
- Towards an Explainable Comparison and Alignment of Feature Embeddings (ICML 2025 · similarity 0.83)
- Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective (NeurIPS 2024 · similarity 0.83)
- Be More Diverse than the Most Diverse: Optimal Mixtures of Generative Models via Mixture-UCB Bandit Algorithms (ICLR 2025 · similarity 0.83)
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.