Research map / Image generation / Editing and style transfer
Edit One for All: Interactive Batch Image Editing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes interactive batch image editing: an edit made on one example image in StyleGAN2 latent space is converted into a globally consistent direction and transferred to other images with per-image closed-form strengths so all reach the same final state.
Problem
Editing many images so that each reaches the same final attribute state currently requires annotating and optimizing each image separately; prior edit-transfer methods do not control edit strength per image.
Approach
Optimize a globally consistent latent direction that reproduces the example edit (image reconstruction loss plus a hyperplane constraint placing the edited example at distance zero), then set each test image's edit strength analytically as its projected distance to that hyperplane, so every image ends at the same attribute value.
Key results
- On keypoint dragging, the method reaches a keypoint distance of 9.467 versus 4.165 for DragGAN, with FID 9.35 versus 9.28, while needing one annotated image instead of 1000.
- Transferring a frontal-face edit to 1000 test images yields yaw MAE of 2.12, compared with 8.14 for DragGAN + GANgealing and 11.29 for random originals.
- Editing 1000 images takes about 82 s total (2 s example edit, 30 s direction optimization, 50 s transfer) versus about 33 minutes for DragGAN on 1000 images.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Optimize the editing direction instead of using the naïve direction (ablation) | Ours · Face pose rotation | yaw-direction correlation R²: 0.744 -> 0.846 | Helped |
| Use the naïve direction rather than the optimized direction (ablation) | Ours · Face pose rotation | yaw-direction correlation R²: 0.846 -> 0.744 | Hurt |
| Remove per-image adjustment of editing strength (ablation) | Ours · Face pose rotation | frontal pose consistency | Hurt |
| Use the optimized direction without dynamic scaling (ablation) | Ours · Face pose rotation | frontal pose | Mixed |
| Transfer an edit with DragGAN + GANgealing instead of the proposed batch method | Face pose rotation · 1000 test faces | Yaw MAE: 2.120 ± 1.818 vs. 8.141 ± 7.221 | Hurt |
| Transfer the edit automatically with the proposed method rather than editing each image with DragGAN | Facial point dragging · Facial point dragging | keypoint distance: 4.165 -> 9.467 | Hurt |
| Transfer the edit automatically with the proposed method rather than editing each image with DragGAN | Facial point dragging · Facial point dragging | FID: 9.28 -> 9.35 | No clear effect |
Surprising findings
- Naive transfer of the example's latent difference does not reliably reproduce the same attribute change on new images, motivating the optimized direction.
Limitations
- Limited to StyleGAN models; does not extend to diffusion-based editors.
- Quantitative evaluation focuses on facial keypoints and yaw; other domains are judged qualitatively.
Experimental setup
Datasets: FFHQ AFHQ Cats AFHQ Dogs MetFaces Human bodies
Benchmarks: FFHQ keypoint dragging (dlib keypoints) yaw pose consistency on 1000 test faces
Metrics: keypoint Euclidean distance FID time per image annotations per edit mean absolute error of yaw
Baselines: DragGAN UserControllableLT DragDiffusion StyleCLIP InstructPix2Pix DragGAN + GANgealing
Builds on: StyleGAN2 DragGAN LARGE GANgealing
Closest papers by meaning
- UltraEdit: Instruction-based Fine-Grained Image Editing at Scale (NeurIPS 2024 · similarity 0.83)
- The Devil is in the Details: StyleFeatureEditor for Detail-Rich StyleGAN Inversion and High Quality Image Editing (CVPR 2024 · similarity 0.83)
- ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.83)
- Zero-shot Image Editing with Reference Imitation (NeurIPS 2024 · similarity 0.83)
- AnyEdit: Mastering Unified High-Quality Image Editing for Any Idea (CVPR 2025 · similarity 0.83)
- UniHuman: A Unified Model For Editing Human Images in the Wild (CVPR 2024 · similarity 0.82)
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.