Research map / Probabilistic and scientific modeling / Variational and generative sampling
Diamond Maps: Efficient Reward Alignment via Stochastic Flow Maps
Peter Holderrieth, Douglas Chen, Luca Eyring, Ishin Shah, Giri Anantharaman, Yutong He, Zeynep Akata, Tommi Jaakkola, Nicholas Matthew Boffi, Max Simchowitz
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 |
|---|---|---|---|
| add Monte Carlo particles (mc=64) to Diamond Maps | Weighted Diamond Maps (FLUX Flow Map) · FLUX GenEval | GenEval: reach 0.785 at mc=64 vs Flow Map Guidance 0.737 and Best-of-N 0.748 | Improved |
| use Weighted Diamond Maps instead of Best-of-N | FLUX Flow Map · UniGenBench++ | UniGenBench++ score: 0.637 -> 0.653 | Improved |
| use Weighted Diamond Maps instead of Flow Map Guidance | FLUX Flow Map · UniGenBench++ | UniGenBench++ score: 0.627 -> 0.653 | Improved |
| add more guidance steps to Flow Map Guidance (ablation) | Flow Map Guidance · FLUX GenEval | GenEval: peak 0.737, after which performance declines with additional steps | Hurt |
| use SNR=1.5 perturbation noise instead of SNR=1.1 (ablation) | Weighted Diamond Maps · FLUX GenEval | GenEval: SNR=1.5 yields the strongest results, insufficient noise (SNR=1.1) limits exploration | Improved |
| use excessive perturbation noise (SNR>=3.0) (ablation) | Weighted Diamond Maps · FLUX GenEval | GenEval: SNR>=3.0 degrades quality, particularly with g=24 | Hurt |
| increase Monte Carlo samples mc (ablation) | Weighted Diamond Maps · FLUX GenEval | GenEval: increasing mc yields consistent gains | Improved |
| use guidance frequency g=10 (ablation) | Weighted Diamond Maps · FLUX GenEval | GenEval: g=10 offers the best trade-off between alignment strength and trajectory coherence | Improved |
| add weighted diamond-map guidance to SANA-Sprint 0.6B | SANA-Sprint 0.6B · GenEval | GenEval mean: 0.70 -> 0.83 | Improved |
| add weighted diamond-map guidance to FLUX Flow Map | FLUX Flow Map · GenEval | GenEval mean: 0.57 -> 0.79 | Improved |
| use Weighted Diamond Maps instead of ReNO on SANA-Sprint | SANA-Sprint 0.6B · GenEval | GenEval mean: 0.81 -> 0.83 with time 30.0 -> 10.0 | Improved |
| use Diamond DDPM early stop sampling instead of iterative denoising and noising (ablation) | Posterior Diamond Maps · posterior sampling | sample quality | Improved |
Closest papers by meaning
- Reward Score Matching: Unifying Reward-based Fine-tuning for Flow and Diffusion Models (NeurIPS 2026 · similarity 0.90)
- $\Psi$-Sampler: Initial Particle Sampling for SMC-Based Inference-Time Reward Alignment in Score Models (NeurIPS 2025 · similarity 0.89)
- VGB for Masked Diffusion Model: Efficient Test-time Scaling for Reward Satisfaction and Sample Editing (NeurIPS 2026 · similarity 0.87)
- Learnable Chernoff Baselines for Provable Inference-Time Alignment (NeurIPS 2026 · similarity 0.87)
- Test-time Alignment of Diffusion Models without Reward Over-optimization (ICLR 2025 · similarity 0.87)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.87)
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.
