Source-linked AI summary
Group Editing: Edit Multiple Images in One Go
Yue Ma, Xinyu Wang, Qianli Ma, Qinghe Wang, Mingzhe Zheng, Xiangpeng Yang, Hao Li, Chongbo Zhao, Jixuan Ying, Harry Yang, Hongyu Liu, Qifeng Chen
TL;DR
GroupEditing targets the difficulty of making coherent edits across related images with varying viewpoints, poses, and layouts. It combines pseudo-video priors with VGGT-based geometric correspondence, alignment-enhanced RoPE, curated data, and a dedicated benchmark. Experiments report superior visual quality, editing consistency, and semantic alignment over existing methods.
Problem
Group-image editing lacks reliable correspondence and consistent appearance and structure across related images, especially in geometrically complex scenes.
Method
GroupEditing treats related images as pseudo-video frames and fuses video-model priors with VGGT geometric correspondences through geometry- and identity-enhanced RoPE.
Results
GroupEditing consistently outperforms state-of-the-art competitors in visual quality, editing consistency, and semantic alignment.
Takeaways & Limitations
The framework provides a training-based approach for consistent multi-image editing supported by GroupEditData and evaluated with GroupEditBench.
Abstract
from arXiv · showhide
In this paper, we tackle the problem of performing consistent and unified modifications across a set of related images. This task is particularly challenging because these images may vary significantly in pose, viewpoint, and spatial layout. Achieving coherent edits requires establishing reliable correspondences across the images, so that modifications can be applied accurately to semantically aligned regions. To address this, we propose GroupEditing, a novel framework that builds both explicit and implicit relationships among images within a group. On the explicit side, we extract geometric correspondences using VGGT, which provides spatial alignment based on visual features. On the implicit side, we reformulate the image group as a pseudo-video and leverage the temporal coherence priors learned by pre-trained video models to capture latent relationships. To effectively fuse these two types of correspondences, we inject the explicit geometric cues from VGGT into the video model through a novel fusion mechanism. To support large-scale training, we construct GroupEditData, a new dataset containing high-quality masks and detailed captions for numerous image groups. Furthermore, to ensure identity preservation during editing, we introduce an alignment-enhanced RoPE module, which improves the model's ability to maintain consistent appearance across multiple images. Finally, we present GroupEditBench, a dedicated benchmark designed to evaluate the effectiveness of group-level image editing. Extensive experiments demonstrate that GroupEditing significantly outperforms existing methods in terms of visual quality, cross-view consistency, and semantic alignment.
1 HKUST 2 THU 3 SJTU 4 University of Technology Sydney
GroupEditing addresses the challenge of editing related images consistently despite variation in viewpoint, pose, and composition. It combines pseudo-video modeling, geometric correspondence, identity preservation, and curated training data to support coherent group-image editing.
- Motivation: Group-image editing must preserve appearance and structure across related images, unlike single-image editing, but prior approaches often produce inconsistent results.The challenge is especially difficult in geometrically complex scenes and when only limited training pairs or image sets are available.
- Method: VGGT supplies explicit dense correspondences that are fused into the video model through geometry-enhanced RoPE, while Identity-RoPE supports consistent identity preservation.The framework combines implicit video priors with explicit geometric cues because semantic correspondence alone can fail under rotation, deformation, or occlusion.
- Method: GroupEditing reformulates related images as pseudo-video frames, using video-model priors for temporal and geometric coherence.The framework treats the image set jointly rather than editing each image independently.
- Evaluation: GroupEditing is evaluated against prior state-of-the-art methods on GroupEditBench and achieves superior visual quality, editing consistency, and semantic alignment.The benchmark contains more than 800 image groups spanning diverse content and styles, with separate local and global editing settings.
- Data: GroupEditData is built through image generation, quality evaluation, segmentation, and captioning to provide high-quality training pairs with precise masks and detailed descriptions.The pipeline filters generated groups for semantic consistency and aesthetic quality; 7,517 high-quality groups form the final training set.
4. Experiment
Experiments evaluate GroupEditing through applications, qualitative comparisons, benchmark testing, and component ablations. The method achieves consistent group editing and leads both automatic metrics and human preference on GroupEditBench.
- Applications: The edited outputs support customized image generation through DreamBooth and LoRA-based fine-tuning.DreamBooth fine-tuning uses 600 steps and enables novel concept generation and editing from the edited images.
- Qualitative comparison: Qualitative comparisons show that GroupEditing edits image groups consistently, while Anydoor, OminiControl, and Edicho struggle with coherent multi-image editing.The comparison covers recent open-sourced methods for local image editing.
- Benchmark evaluation: GroupEditBench contains 800 image sets spanning diverse content and styles, with separate local and global editing evaluations.Methods are evaluated using editing consistency, aesthetic quality, and human rankings across identity consistency, aesthetic, appearance fidelity, and overall quality.
- Benchmark evaluation: GroupEditing achieves the top result in both automatic metrics and human preference on the collected benchmark.The benchmark uses feature similarity for editing consistency, LAION-Aesthetic Score for visual quality, and average volunteer rankings, where lower is better.
- Ablation study: Geometry-enhanced RoPE improves spatial awareness and geometric transformation accuracy in ablations.The paper reports both visual improvements and quantitative validation in the ablation study.
- Ablation study: Identity-RoPE produces more consistent and stable object identity across different images.The reported visual and quantitative ablations attribute improved identity consistency to incorporating Identity-RoPE layers.
5. Conclusion
GroupEditing combines geometric correspondences, video-model priors, and alignment-enhanced identity preservation for consistent multi-image editing. Evaluations and visual comparisons report strong consistency, edit quality, and performance against state-of-the-art methods.
- GroupEditing combines explicit geometric correspondences from VGGT with implicit priors from pre-trained video models.The framework also uses alignment-enhanced RoPE to preserve identity across diverse images.
- The framework is supported by GroupEditData, which provides high-quality image groups, precise segmentation masks, and detailed captions for training.
- GroupEditBench evaluates group-level editing, while visual comparisons cover both local and global image editing.The benchmark comparisons include state-of-the-art image-editing methods.
- GroupEditing outperforms state-of-the-art methods in visual quality, editing consistency, and semantic alignment.The paper presents this outcome as evidence for broad multi-image editing applications.
A. Rationale
Compiling the supplementary material with the main paper allows cross-references in both directions and includes the supplementary pages with the arXiv submission.
- The supplementary material can refer back to sections of the main paper.
- The main paper can explicitly refer forward to subsections within the supplementary material.
- The supplementary material is included at the end of the main paper when submitted to arXiv.
B. Realted Work
The related-work discussion places image editing within diffusion-based generative modeling and distinguishes inference-time zero-shot methods from training-based approaches.
- Diffusion-based generative models have driven substantial progress in image editing.
- Existing image-editing techniques are broadly divided into inference-time zero-shot methods and training-based methods.
- The paper relates group editing to the coherence and geometric priors learned by large-scale video models.
D. Demo
The paper provides a demo video and project page to illustrate the proposed method and its results.
- A demo video is provided for illustrating the method.
- The paper also provides a project page containing the demonstration materials.
- The authors recommend viewing the demo for better illustration.