Research map / Image generation / Editing and style transfer
Training-Free Reward-Guided Image Editing via Trajectory Optimal Control
Jinho Chang, Jaemin Kim, Jong Chul Ye
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 |
|---|---|---|---|
| optimize entire trajectory via adjoint guidance instead of inversion + posterior-mean guidance (TFG) | Ours · Human preference, SD1.5 | ImageReward / LPIPS: ImageReward 1.7053 -> 1.8914; LPIPS 0.2927 -> 0.1717 | Improved |
| use trajectory control instead of gradient ascent | Ours · Human preference, SD1.5 | HPSv2: 0.2247 -> 0.2526 | Improved |
| compare against gradient ascent on target reward | Ours · Human preference, SD1.5 | ImageReward: 1.9088 -> 1.8914 | Hurt |
| trajectory control vs FreeDoM in style transfer | Ours · Style transfer, SD1.5 | DINO style similarity: 0.2250 -> 0.2467 | Improved |
| trajectory control vs gradient ascent in counterfactual generation | Ours · Counterfactual, ImageNet-1k, SD1.5 | CLIPScore: logit 24.875 -> 23.372; CLIPScore 0.1908 -> 0.1936 | Mixed |
| trajectory control vs FreeDoM in text-guided editing | Ours · CelebA-HQ, SD1.5 | ImageReward: -0.1100 -> 0.0976 | Improved |
| same method on flow-matching StableDiffusion 3 | Ours · Human preference, SD3 | ImageReward: 1.5887 -> 1.8529 vs best inversion baseline | Improved |
| use noiseless reverse path instead of Markovian stochastic initial trajectory (ablation) | Ours · Initial trajectory choice | stability to hyperparameters | Improved |
| increase compute for baselines (more N_recur and N_iter) (ablation) | Inversion-based baselines · Human preference | reward | Hurt |
| trajectory control vs baselines in user study | Ours · User study | reward alignment rating: best baseline 3.28 -> 3.67 | Improved |
Closest papers by meaning
- Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL (NeurIPS 2026 · similarity 0.89)
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.89)
- AutoEdit: Automatic Hyperparameter Tuning for Image Editing (NeurIPS 2025 · similarity 0.89)
- h-Edit: Effective and Flexible Diffusion-Based Editing via Doob's h-Transform (CVPR 2025 · similarity 0.89)
- High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.88)
- SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing (NeurIPS 2025 · similarity 0.88)
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.
