Research map / Image generation / Fast diffusion sampling
NoiseTilt: Noise-Tilted Reverse Kernels for Diffusion Reward Alignment
Jisung Hwang, Yunhong Min, Jaihoon Kim, I-Chao Shen, 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 |
|---|---|---|---|
| NoiseTilt (25 NFE) instead of SVDD (500 NFE) | FLUX · Aesthetic image generation (target Aesthetic Score) | Aesthetic Score: 7.1363 -> 7.4510 | Improved |
| add Best-of-N to NoiseTilt (500 NFE) (ablation) | NoiseTilt · Aesthetic image generation (target Aesthetic Score) | Aesthetic Score: 7.4510 -> 7.9656 | Improved |
| add Best-of-N to NoiseTilt (500 NFE) (ablation) | NoiseTilt · Aesthetic image generation (target Aesthetic Score) | ImageReward: 1.2565 -> 1.1669 | Hurt |
| add Best-of-N to NoiseTilt (500 NFE) (ablation) | NoiseTilt · Text-aligned image generation (target PickScore) | PickScore: 0.2224 -> 0.2327 | Improved |
| add Best-of-N to NoiseTilt (500 NFE) (ablation) | NoiseTilt · Text-aligned image generation (target PickScore) | VQA Score: 0.8210 -> 0.8017 | Hurt |
| NoiseTilt (25 NFE) instead of SVDD (500 NFE) | FLUX · Text-aligned image generation (target PickScore) | PickScore: 0.2204 -> 0.2224 | Improved |
| NoiseTilt instead of DPS | Wan2.1 · Preference-aligned video generation (VBench) | VideoReward: -0.130 -> 3.465 | Improved |
| NoiseTilt instead of Base | Wan2.1 · Preference-aligned video generation (VBench) | VideoReward: -0.399 -> 3.465 | Improved |
| NoiseTilt instead of FreeDoM | Wan2.1 · Preference-aligned video generation (VBench) | Dynamic Degree: 0.6900 -> 0.9500 | Improved |
| DPS guidance instead of Base | Wan2.1 · Preference-aligned video generation (VBench) | Aesthetic Quality: 0.6104 -> 0.5867 | Hurt |
| add MixGRPO fine-tuning | FLUX · Text-aligned image generation with fine-tuned model | PickScore: 0.2054 -> 0.2166 | Improved |
| add NoiseTilt on top of MixGRPO instead of DPS | MixGRPO · Text-aligned image generation with fine-tuned model | PickScore: 0.2235 -> 0.2545 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- MixGRPO — MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE (ECCV 2026) · better on 1 of 1 shared result
- FreeDoM — FreeDoM: Training-Free Energy-Guided Conditional Diffusion Model (ICCV 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Inference-Time Alignment of Diffusion Models with Direct Noise Optimization (ICML 2025 · similarity 0.91)
- VGB for Masked Diffusion Model: Efficient Test-time Scaling for Reward Satisfaction and Sample Editing (NeurIPS 2026 · similarity 0.88)
- Test-time Alignment of Diffusion Models without Reward Over-optimization (ICLR 2025 · similarity 0.88)
- Training-Free Diffusion Model Alignment with Sampling Demons (ICLR 2025 · similarity 0.88)
- Hierarchical Variational Policies for Reward-Guided Diffusion (NeurIPS 2026 · similarity 0.88)
- Reward Score Matching: Unifying Reward-based Fine-tuning for Flow and Diffusion Models (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.
