Constrained Look-ahead Guidance for Interference-Aware Flow Editing
Organizations: ELLIS Institute Finland · Aalto University
Abstract
Diffusion and flow models enable powerful generative modeling, but controlled editing remains challenging due to the need to balance source fidelity and edit compliance along the generative trajectory. We propose a principled, inversion-free framework for flow-based image editing from a trajectory centric perspective. Rather than relying on inversion or heuristic guidance, we introduce a one-step look-ahead mechanism that evaluates the effect of candidate updates on both edit and fidelity rewards estimated via denoising, and adjusts the current update based on this evaluation. At each step, our method applies reward-driven corrections guided by the look-ahead evaluation under an explicit fidelity constraint, resulting in stable and coherent trajectory evolution. This look-ahead design mitigates gradient interference between edit and fidelity objectives by anticipating their interaction in future states, leading to improved trade-offs. Experiments demonstrate that the proposed approach achieves stronger semantic edits while better preserving source content compared to existing strong baselines.
Figures & tables
| Method | Structure | Background Preservation | CLIP Similarity | ||||
|---|---|---|---|---|---|---|---|
| PSNR | LPIPS | MSE | SSIM | Whole | Edited | ||
| DDIM Inversion [ 41 ] | 0.1229 | 12.7867 | 0.3578 | 0.0763 | 0.6025 | 26.7718 | 23.9214 |
| TABA [ 44 ] | 0.1233 | 12.8892 | 0.3496 | 0.0744 | 0.6056 | 27.4278 | 24.2488 |
| RF-Inversion [ 38 ] | 0.0561 | 19.1570 | 0.2290 | 0.0171 | 0.6600 | 26.1720 | 23.3210 |
| RF-Solver-Edit [ 48 ] | 0.0249 | 24.4093 | 0.1149 | 0.0057 | 0.8446 | 25.6668 | 22.7751 |
| FlowChef-Edit [ 35 ] | 0.0395 | 24.3100 | 0.1239 | – | 0.7960 | 25.3200 | 22.3400 |
| Method | Runtime / image (s) | Peak memory (GiB) |
|---|---|---|
| FlowEdit [ 22 ] | 7.00 | 34.18 |
| TABA [ 44 ] | 7.80 | 32.05 |
| DDIM Inversion [ 41 ] | 8.08 | 32.05 |
| RF-Solver-Edit [ 48 ] | 16.45 | 32.59 |
| RF-Inversion [ 38 ] | 25.90 | 35.72 |
| ITOC [ 3 ] | 139.25 | 40.93 |
| Update rule | Structure | PSNR | LPIPS | MSE | SSIM | CLIP Similarity (Whole) | CLIP Similarity (Edited) |
|---|---|---|---|---|---|---|---|
| Edit-first | 0.0138 | 26.2152 | 0.0841 | 0.0033 | 0.8258 | 26.9845 | 22.0966 |
| Look-ahead | 0.0091 | 27.2950 | 0.0751 | 0.0026 | 0.8476 | 27.2370 | 22.0130 |
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | LPIPS | DINO | DreamSim | CLIP-T |
|---|---|---|---|---|
| FlowEdit [ 22 ] | 0.1854 | 0.7402 | 0.2576 | 29.2862 |
| Ours | 0.1573 | 0.8443 | 0.1144 | 29.6528 |
| Method | Structure | PSNR | LPIPS | MSE | SSIM | CLIP Similarity (Whole) | CLIP Similarity (Edited) | Runtime / image (s) | Peak GPU memory (GiB) |
|---|---|---|---|---|---|---|---|---|---|
| Look-ahead (Ours) | 0.0091 | 27.2950 | 0.0751 | 0.0026 | 0.8476 | 27.2370 | 22.0130 | 185.88 | 55.82 |
| Look-ahead with one-step denoising | 0.0114 | 27.3860 | 0.0757 | 0.0025 | 0.8510 | 27.1912 | 22.0686 | 243.70 | 64.88 |
| Update rule | Structure | PSNR | LPIPS | MSE | SSIM | CLIP Similarity (Whole) | CLIP Similarity (Edited) |
|---|---|---|---|---|---|---|---|
| Edit-first | 0.0175 | 25.0160 | 0.0965 | 0.0045 | 0.8442 | 28.7452 | 23.7910 |
| Look-ahead | 0.0176 | 25.0050 | 0.0957 | 0.0045 | 0.8433 | 28.6262 | 23.7916 |
| Method | Edit Compliance | Source Fidelity |
|---|---|---|
| DDIM Inversion | 2.83 | 2.00 |
| TABA | 2.92 | 2.21 |
| RF-Inversion | 3.12 | 2.58 |
| RF-Solver-Edit | 3.33 | 3.29 |
| FlowEdit | 3.46 | 3.21 |
| FlowAlign | 3.04 | 2.83 |
| Sweep | Value | Structure | PSNR | LPIPS | MSE | SSIM | CLIP Whole | CLIP Edited | Violations/image | Updates/image |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 0.0126 | 27.4450 | 0.0795 | 0.0025 | 0.8454 | 25.9090 | 21.1890 | – | 28.00 | |
| 2 | 0.0101 | 27.4370 | 0.0770 | 0.0025 | 0.8522 | 25.7660 | 21.2480 | – | 56.00 | |
| 3 | 0.0091 | 27.2950 | 0.0751 | 0.0026 | 0.8476 | 27.2370 | 22.0130 | – | 84.00 | |
| 4 | 0.0103 | 27.4060 | 0.0752 | 0.0024 | 0.8508 | 27.1870 | 22.1060 | – | 112.00 | |
| 1 | 0.0129 | 27.2300 | 0.0835 | 0.0026 | 0.8409 | 26.4750 | 21.1020 | 64.22 | – | |
| 2 | 0.0101 | 27.2550 | 0.0781 | 0.0025 | 0.8456 | 27.9370 | 21.6040 | 55.61 | 103.56 |
| Sweep | Value | Structure | PSNR | LPIPS | MSE | SSIM | CLIP Whole | CLIP Edited |
|---|---|---|---|---|---|---|---|---|
| 0.25 | 0.0094 | 27.3820 | 0.0769 | 0.0025 | 0.8498 | 26.9050 | 21.9750 | |
| 0.5 | 0.0094 | 27.2030 | 0.0745 | 0.0026 | 0.8471 | 27.1840 | 21.7970 | |
| 1.0 | 0.0091 | 27.2950 | 0.0751 | 0.0026 | 0.8476 | 27.2370 | 22.0130 | |
| 2.0 | 0.0105 | 27.3680 | 0.0771 | 0.0025 | 0.8519 | 26.7390 | 22.4200 | |
| 4.0 | 0.0103 | 27.5620 | 0.0710 | 0.0024 | 0.8518 | 26.5090 | 21.2290 | |
| 1.0 | 0.0095 | 27.2720 | 0.0766 | 0.0026 | 0.8508 | 26.8310 | 21.8640 |
| Type | Variant | Structure | PSNR | LPIPS | MSE | SSIM | Source CLIP | Target CLIP | CLIP Edited | Margin |
|---|---|---|---|---|---|---|---|---|---|---|
| Edit reward | Text-image | 0.0100 | 27.4210 | 0.0765 | 0.0025 | 0.8518 | 34.4100 | 37.1330 | 26.8110 | 2.7230 |
| Directional | 0.0094 | 27.5340 | 0.0749 | 0.0025 | 0.8507 | 22.3930 | 27.2300 | 22.4150 | 4.8370 | |
| Contrastive | 0.0091 | 27.2950 | 0.0751 | 0.0026 | 0.8476 | 18.7140 | 27.2370 | 22.0130 | 8.5240 | |
| Initialization | Inversion-free | 0.0091 | 27.2950 | 0.0751 | 0.0026 | 0.8476 | – | 27.2370 | 22.0130 | – |
| Inversion | 0.0155 | 25.5560 | 0.0845 | 0.0038 | 0.8385 | – | 27.5780 | 21.9290 | – |
| Reward | Structure | PSNR | LPIPS | MSE | SSIM | CLIP Whole | CLIP Edited |
|---|---|---|---|---|---|---|---|
| CLIP-contrastive | 0.0091 | 27.2950 | 0.0751 | 0.00260 | 0.8476 | 27.2370 | 22.0130 |
| ImageReward | 0.0125 | 27.1130 | 0.0777 | 0.00256 | 0.8475 | 26.1243 | 21.9768 |
| HPS v2.1 | 0.0124 | 26.9636 | 0.0776 | 0.00274 | 0.8471 | 27.1997 | 22.4951 |
| Objective | Weight | Structure | PSNR | LPIPS | MSE | SSIM | CLIP Whole | CLIP Edited |
|---|---|---|---|---|---|---|---|---|
| Latent- | 0 | 0.0091 | 27.2950 | 0.0751 | 0.0026 | 0.8476 | 27.2370 | 22.0130 |
| DINO | – | 0.2106 | 9.6213 | 0.4456 | 0.1427 | 0.5122 | 26.8204 | 22.7671 |
| Latent- + DINO | 0.005 | 0.0118 | 26.9860 | 0.0862 | 0.0028 | 0.8478 | 26.5615 | 21.2238 |
| 0.020 | 0.0096 | 27.4742 | 0.0736 | 0.0024 | 0.8548 | 26.6459 | 20.6478 | |
| 0.050 | 0.0173 | 26.6272 | 0.0952 | 0.0036 | 0.8392 | 25.7457 | 21.4099 | |
| 0.100 | 0.0154 | 27.0727 | 0.0851 | 0.0029 | 0.8453 | 26.9130 | 21.3573 |