Research map / Image generation / Fast diffusion sampling
Calibrating Generative Models to Distributional Constraints
Henry Smith, Nathaniel Lee Diamant, Brian L. Trippe
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 |
|---|---|---|---|
| calibrate with CGM-relax | Genie2 · CATH secondary structure | symmetrized KL to CATH: nearly fivefold improvement | Improved |
| calibrate with CGM-relax | ESM3-open · CATH secondary structure | symmetrized KL to CATH: twofold improvement | Improved |
| calibrate with CGM-relax | Genie2 / ESM3-open · protein design | design failures | Hurt |
| calibrate with CGM-reward (15 cutoff pairs) | Genie2 · CATH secondary structure | symmetrized KL: two times (Genie2), 1.6 times (ESM3-open) | Improved |
| increase Genie2 sampling noise from 0.5 to 1 instead of calibration | Genie2 · protein design | design failure rate: 74% vs 14% | Hurt |
| increase ESM3 sampling temperature from 0.7 to 1 | ESM3-open · protein design | design failure rate: 97% | Hurt |
| calibrate with CGM-relax | conditional TarFlow · AFHQ wildlife | total variation to uniform: 0.306 -> 0.101 | Improved |
| calibrate with CGM-relax | conditional TarFlow · AFHQ wildlife | FID: 15.9 -> 21.0 | Hurt |
| calibrate with CGM-reward | conditional TarFlow · AFHQ wildlife | constraint violation | No clear effect |
| calibrate with CGM-reward | Gemma-2-9B-IT · gender-profession imbalance | miscalibration | Improved |
| calibrate with CGM-relax | Gemma-2-9B-IT · gender-profession imbalance | gender imbalance: more than threefold reduction | Improved |
| calibrate on explicit professions, test on held-out professions | Gemma-2-9B-IT · six held-out professions | gender imbalance: four improved, two worse | Mixed |
Closest papers by meaning
- Calibrating Generative Models to Feature Distributions with MMD Finetuning (NeurIPS 2026 · similarity 0.88)
- Bridging Model-Based Optimization and Generative Modeling via Conservative Fine-Tuning of Diffusion Models (NeurIPS 2024 · similarity 0.87)
- Aligning Few-Step Generative Model via Amortizing Sample-Based Variational Inference (NeurIPS 2026 · similarity 0.86)
- Conformal Generative Modeling with Improved Sample Efficiency through Sequential Greedy Filtering (ICLR 2025 · similarity 0.86)
- Score-based generative models are provably robust: an uncertainty quantification perspective (NeurIPS 2024 · similarity 0.86)
- Algorithm- and Data-Dependent Generalization Bounds for Diffusion Models (NeurIPS 2025 · similarity 0.86)
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.
