Research map / Image generation / Fast diffusion sampling
Gradient Preconditioning for Efficient and Reliable Reward-Guided Generation
Jisung Hwang, Minhyuk Sung
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 |
|---|---|---|---|
| project reward gradients onto the white Gaussian noise feasible set instead of regularizing the latent | test-time latent optimization with FLUX-schnell · Aesthetic Score optimization, 200 iterations | Aesthetic Score (target): 7.133 -> 8.908 (vs MPGR) | Improved |
| use gradient preconditioning with only 60 iterations | gradient preconditioning · Aesthetic Score optimization, FLUX | wall-clock time (s) at matched Aesthetic Score: 235.5 -> 69.7 | Improved |
| use gradient preconditioning instead of regularization | FLUX-schnell · held-out PickScore under Aesthetic Score optimization | PickScore (held-out): 0.219 -> 0.220 | No clear effect |
| add gradient preconditioning versus unoptimized sampling | FLUX-schnell · diversity under Aesthetic Score optimization | Inception Score / Vendi Score: IS 22.33 -> 21.10; Vendi 6.42 -> 6.97 | No clear effect |
| add projection step to every iteration | gradient preconditioning · FLUX runtime | share of runtime: 0.04% of runtime | No clear effect |
| tune the regularization weight lambda for baselines (ablation) | regularization-based baselines · reward-guided generation, FLUX | reward trade-off | No clear effect |
| use soft spatial-correlation regularizers (PRNO, MPGR) to avoid reward hacking | baseline regularizers · qualitative FLUX results | realism of outputs | Mixed |
Closest papers by meaning
- Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL (NeurIPS 2026 · similarity 0.88)
- Reward-Guided Iterative Refinement in Diffusion Models at Test-Time with Applications to Protein and DNA Design (ICML 2025 · similarity 0.87)
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.87)
- Exact risk curves of signSGD in High-Dimensions: quantifying preconditioning and noise-compression effects (ICML 2025 · similarity 0.86)
- VGB for Masked Diffusion Model: Efficient Test-time Scaling for Reward Satisfaction and Sample Editing (NeurIPS 2026 · similarity 0.86)
- Constrained Stochastic Spectral Preconditioning Converges for Nonconvex Objectives (NeurIPS 2026 · 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.
