Skip to content

Group Editing: Edit Multiple Images in One Go

Conference: CVPR 2026 arXiv: 2603.22883 Code: https://group-editing.github.io/ Area: Diffusion Models / Image Editing Keywords: consistent multi-image editing, video diffusion prior, geometric correspondence, RoPE positional encoding, pseudo-video

TL;DR

This paper proposes GroupEditing, which reconstructs a group of related images as pseudo-video frames and combines explicit geometric correspondences from VGGT with the implicit temporal prior of a video diffusion model. Two specially designed positional encodings—Ge-RoPE and Identity-RoPE—are introduced to inject correspondence information, enabling cross-view consistent group image editing that significantly outperforms existing methods in visual quality, editing consistency, and semantic alignment.

Background & Motivation

  1. Background: Existing image editing methods (e.g., InstructPix2Pix, ControlNet) focus primarily on single-image editing. In virtual content creation, digital commerce, and similar applications, users frequently need to apply consistent modifications across multi-view images of the same subject—such as uniformly recoloring a digital character's clothing or stylizing product images from multiple angles.
  2. Limitations of Prior Work: Editing images one by one leads to appearance and structural inconsistencies. Optimization-based propagation methods (e.g., editing one image and then propagating the changes) suffer from poor generalization and tend to produce artifacts. Optimization-free methods (e.g., Edicho) rely on semantic correspondence and tracking tools and can only handle a small number of images.
  3. Key Challenge: In geometrically complex scenarios (e.g., target rotation, occlusion, deformation), semantic matching based solely on attention features is insufficiently precise. Identifying "the left eye across different viewpoints" or "tracking a logo on a T-shirt rotated by 30°" poses significant challenges for existing methods.
  4. Goal: To establish reliable cross-image correspondences within a geometrically diverse group of related images, enabling multi-image consistent editing from a single instruction.
  5. Key Insight: The authors identify two key observations: (1) Implicit correspondence—video models inherently possess a temporal consistency prior, which can be inherited by treating an image group as a "pseudo-video"; (2) Explicit correspondence—the implicit correspondences of video models alone are insufficient in geometrically complex scenarios, necessitating dense geometric matching from VGGT as a complement.
  6. Core Idea: The multi-image editing problem is reformulated as a pseudo-video generation problem, fusing explicit geometric correspondences (VGGT) with an implicit temporal prior (video diffusion model) by injecting correspondence information through specially designed positional encodings.

Method

Overall Architecture

The inputs consist of a group of related images along with their corresponding segmentation masks and a text editing instruction. Images are first encoded into latent space via a VAE encoder and arranged as a pseudo-video sequence along the temporal dimension. Within the Transformer backbone of the WAN-2.1 video diffusion model, two enhanced RoPE positional encodings are injected: Ge-RoPE for cross-view geometric alignment and Identity-RoPE for intra-image identity preservation of the target. Explicit geometric feature tokens extracted by VGGT are concatenated to the latent token sequence and participate in self-attention computation. The edited multi-view consistent images are then produced by the decoder.

Key Designs

  1. Data Construction Pipeline (GroupEditData):

    • Function: Constructs a large-scale multi-image editing training dataset.
    • Mechanism: Gemini 2.5 is used to generate image groups (18,248 groups) from manually written text instructions. SAM and Grounding DINO are applied for target segmentation to obtain masks. Qwen-VL-Max is then employed for consistency and aesthetic evaluation to filter for quality, retaining 7,517 high-quality groups. Each group contains images, masks, full-image descriptions, and segmented region descriptions.
    • Design Motivation: Large-scale paired multi-image editing data is currently unavailable. This pipeline provides the critical infrastructure that makes training feasible.

  2. Geometry-enhanced RoPE (Ge-RoPE):

    • Function: Injects explicit geometric correspondence information extracted by VGGT into the positional encodings to achieve fine-grained spatial alignment across viewpoints.
    • Mechanism: A pixel-level displacement field \(\Delta(h,w) = (\Delta_h, \Delta_w)\) is obtained from VGGT, scaled to the latent-space resolution, and smoothed with a Gaussian kernel (\(\mu=21, \sigma=11\)) while prioritizing high-confidence correspondences. The smoothed displacements are added to the original spatial grid indices to construct a warped grid \(\tilde{h} = h + \Delta_h^{\text{smooth}}\), and nearest-neighbor indexing into a precomputed frequency bank generates the geometry-aware RoPE encoding.
    • Design Motivation: The implicit correspondences of video models are insufficiently accurate in geometrically complex scenes. Ge-RoPE uses explicit displacement fields to inform the model which position in image B corresponds to position \((h,w)\) in image A, substantially improving spatial alignment precision.

  3. Identity-RoPE:

    • Function: Ensures identity consistency of the same target across different images.
    • Mechanism: The minimum bounding rectangle \(\mathcal{R}_t\) of the target in each image is identified via segmentation masks. Pixel coordinates within the rectangle are normalized to local coordinates relative to the rectangle's origin: \((\tilde{h}, \tilde{w}) = (h - y_1^{(t)}, w - x_1^{(t)})\). Consequently, the same target region across different images receives identical positional encodings, regardless of its absolute position within each image.
    • Design Motivation: Targets may appear at different absolute positions across viewpoints, causing standard positional encodings to treat them as distinct entities. Identity-RoPE enables all instances of "the cat's face" across images to share the same positional signal through coordinate normalization, thereby preserving identity consistency.

Loss & Training

Training is conducted on WAN-2.1 (a Transformer-based video diffusion model) using the AdamW optimizer (weight decay 0.01, learning rate \(1 \times 10^{-4}\)), at a resolution of \(528 \times 528\), with batch size 8 on 8 A800 GPUs. The training objective is the standard velocity-field prediction loss.

Key Experimental Results

Main Results

Method CLIP-Score↑ Aesthetic↑ DINO-Score↑ Edit Consistency↑ PSNR↑
Anydoor 0.2728 4.72 0.7208 0.8697 0.6182
OminiControl 0.2902 5.10 0.7326 0.8676 0.6457
Edicho 0.3059 4.89 0.8080 0.8988 0.6935
GroupEditing 0.3122 5.39 0.8168 0.9239 0.7624

User study results (ranked 1=best to 4=worst): GroupEditing ranks first across all four dimensions—identity consistency (1.67), aesthetics (1.46), appearance fidelity (1.50), and overall quality (1.47).

Ablation Study

Configuration CLIP-Score↑ Aesthetic↑ DINO-Score↑ Edit Consistency↑
w/o VGGT 0.2728 4.72 0.7208 0.8616
w/o Ge-RoPE 0.2902 4.89 0.7326 0.8697
w/o Identity-RoPE 0.2902 4.89 0.7326 0.9108
Full model 0.3122 5.39 0.8168 0.9239

Key Findings

  • Explicit geometric features from VGGT contribute the most: removing them causes DINO-Score to drop from 0.8168 to 0.7208 and edit consistency to drop from 0.9239 to 0.8616.
  • Identity-RoPE primarily improves edit consistency (0.9108→0.9239), with a comparatively smaller contribution to visual quality.
  • Edited results can be directly applied to DreamBooth/LoRA personalization and Must3R 3D reconstruction, validating cross-view consistency.

Highlights & Insights

  • The pseudo-video reformulation is highly elegant: Recasting multi-image editing as video editing inherits the temporal consistency prior of video models at no additional cost—a clean and effective problem transformation.
  • The fusion of explicit and implicit correspondences: Ge-RoPE injects geometric information via positional encodings rather than modifying attention weights, yielding a lightweight yet effective integration mechanism.
  • Engineering value of the data construction pipeline: The fully automated text→generation→filtering→annotation pipeline is transferable to other tasks that require paired training data.

Limitations & Future Work

  • Training data is generated by Gemini rather than sourced from real multi-view images, which may limit generalization to real-world scenarios.
  • Editing quality depends on the accuracy of VGGT's geometric correspondences and may degrade when VGGT estimates are unreliable.
  • Resolution is fixed at \(528 \times 528\); extension to high-resolution settings has not been validated.
  • Segmentation masks are required as input, which increases the barrier to practical use.
  • vs. Edicho: Edicho performs zero-shot consistent editing via semantic correspondence and tracking tools, but is constrained to a small number of images. GroupEditing is the first training-based framework, scaling to larger image groups through a combination of data and model design.
  • vs. Frame2Frame/ChronoEdit: These methods leverage video models to enhance temporal consistency in single-image editing. GroupEditing extends this paradigm by treating multiple images as a unified pseudo-video.
  • vs. ControlNet/T2I-Adapter: These are general single-image conditional control methods. GroupEditing focuses specifically on consistency constraints across multiple images.

Rating

  • Novelty: ⭐⭐⭐⭐ The combination of pseudo-video reformulation and dual-RoPE injection is creative, though individual components are not entirely novel.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Covers quantitative evaluation, qualitative results, user study, ablation study, and downstream application validation—relatively comprehensive.
  • Writing Quality: ⭐⭐⭐⭐ Logically clear with rich illustrations.
  • Value: ⭐⭐⭐⭐ Multi-image consistent editing addresses a practical need; the first training-based framework has pioneering significance.