Group Editing : Edit Multiple Images in One Go

The paper introduces GroupEditing, a novel framework that achieves consistent and unified modifications across diverse image groups by fusing explicit geometric correspondences from VGGT with implicit temporal priors from video models, supported by a new dataset, an identity-preserving alignment module, and a dedicated benchmark.

The Gist

Imagine you are a director filming a movie scene. You have a character (let's say, a cool robot) who needs to walk through a forest, stand in a city, and sit on a beach.

If you ask a standard AI artist to draw the robot in the forest, then draw it again in the city, and then again on the beach, you might get three different robots. One might have blue eyes, the other green; one might be wearing a hat, the other not. They look like siblings, but not the same person.

"Group Editing" is the new tool that solves this problem. It allows you to edit a whole set of related images at once, ensuring that the "character" stays exactly the same, no matter where they are or how the camera moves.

Here is how the paper explains this magic, broken down into simple concepts:

1. The Problem: The "Copy-Paste" Glitch

Current AI tools are great at editing one photo. But if you try to edit a group of photos (like a 360-degree view of a product or a character in different poses), the AI gets confused. It treats every photo as a separate universe.

• The Result: You try to put sunglasses on a dog in four different photos. In photo 1, the sunglasses are on the nose. In photo 2, they are floating in the sky. In photo 3, they are on the tail. It's a mess.

2. The Solution: The "Pseudo-Video" Trick

The researchers came up with a clever idea: What if we pretend these separate photos are actually frames in a video?

• The Analogy: Imagine you have a stack of 10 photos of a cat jumping. If you flip them quickly, it looks like a video.
• The Magic: Video AI models are already very smart at understanding that "Frame 1" and "Frame 2" are the same cat moving through time. They know how the cat's ear moves from left to right.
• The Move: The researchers take a group of static images and trick the AI into thinking, "Oh, this is a video!" This lets the AI use its "video brain" to keep the edits consistent across all the images.

3. The Two Brains: "Implicit" and "Explicit"

To make this work perfectly, the system uses two different types of "glue" to hold the images together:

• Brain A (The Implicit Video Brain): This is the "video model" mentioned above. It uses its general knowledge of how objects move and change shape over time. It's like a human director who intuitively knows, "If the character turns left, their left ear should move right."
• Brain B (The Explicit Geometry Brain): Sometimes, the video brain isn't enough, especially if the images are very different (e.g., a top-down view vs. a side view). So, they bring in a super-precise map-maker called VGGT.
  • The Analogy: If the video brain is a skilled painter, VGGT is a laser scanner. It measures the exact distance between the dog's nose in Photo 1 and the dog's nose in Photo 2. It creates a strict "connect-the-dots" map so the AI knows exactly where to put the sunglasses in every single picture.

4. The Special Glue: "RoPE"

How do they stick these two brains together? They invented a special kind of "positional glue" called RoPE (Rotary Positional Embedding). Think of it as a coordinate system that tells the AI where everything is.

They created two special versions of this glue:

• Geometry-Enhanced RoPE: This uses the laser scanner (VGGT) data to tell the AI, "Hey, even though the dog is tilted in this photo, the nose is here." It fixes the spatial alignment.
• Identity-Enhanced RoPE: This is the "Name Tag" glue. It ensures that the AI remembers, "This is Buster the dog, not just any dog." It locks the dog's features (fur color, eye shape) so they don't change when you edit the background.

5. The Training Data: The "Super-Teacher"

To teach this new AI, they couldn't just use random photos. They needed a massive library of examples where the same object appeared in many different angles, with perfect descriptions.

• They built a pipeline that automatically generated thousands of these "group" examples.
• They used other AIs to draw the images, then used other AIs to check if the images looked good and if the descriptions were accurate.
• The result is a massive dataset called GroupEditData, which acts like a super-teacher, showing the model exactly how to keep things consistent.

6. Why Does This Matter?

This isn't just about making funny pictures. This technology is a game-changer for:

• Virtual Avatars: Making sure your digital twin looks the same in every photo on your social media.
• E-Commerce: Showing a shoe from 10 different angles, and if you change the color to red, all 10 angles turn red perfectly.
• 3D Modeling: If you edit a photo of a building, this tool helps turn those 2D edits into a 3D model without the building looking broken or warped.

Summary

Group Editing is like giving an AI a "group hug." Instead of treating every photo as a stranger, it treats them as a family. By combining the intuitive flow of video AI with the laser-precision of geometric mapping, it ensures that when you edit a group of images, the changes happen perfectly in sync, keeping the identity of the subject intact across the entire set.

Technical Summary

1. Problem Statement

The paper addresses the challenge of Group-Image Editing, which aims to apply consistent and unified modifications across a set of related images (e.g., the same object from different angles, poses, or lighting conditions).

• Core Challenge: Unlike single-image editing, group editing requires maintaining identity integrity and structural coherence across multiple images. Existing methods typically operate on a per-image basis, leading to inconsistencies in appearance, pose, and semantic alignment when applied to groups.
• Limitations of Prior Work:
  • Optimization-based methods: Struggle with generalization, often producing artifacts when propagating edits across diverse geometric variations.
  • Optimization-free methods: Rely on attention features or tracking tools but are limited to small image sets and fail in complex geometric scenarios (e.g., heavy rotation or occlusion).
  • Data Scarcity: There is a lack of high-quality training pairs with precise masks and detailed captions for image groups.

2. Methodology: GroupEditing

The authors propose GroupEditing, a training-based framework that reformulates a set of related images as pseudo-video frames. This allows the model to leverage the strong temporal coherence priors of pre-trained video diffusion models while integrating explicit geometric constraints.

A. Data Curation (GroupEditData)

To enable large-scale training, the authors constructed GroupEditData, a dataset containing over 7,000 image groups.

• Pipeline:
  1. Generation: Uses Gemini 2.5 (T2I) to generate diverse image groups from single text instructions.
  2. Segmentation: Uses Segment Anything (SAM) and Grounding DINO to create precise pixel-level masks.
  3. Quality Control: Employs Qwen-VL-Max for semantic consistency checks and an aesthetic evaluator to filter out low-quality or inconsistent images.
  4. Annotation: Generates detailed captions for the whole image and specific descriptions for masked regions.

B. Model Architecture

The framework is built upon a transformer-based video diffusion model (WAN-2.1) with two novel modules to handle correspondence:

1. Explicit Correspondence (VGGT Integration):

   • The model extracts dense geometric correspondences using VGGT (a visual feature matching model).
   • These explicit geometric cues are injected into the diffusion model to ensure spatial alignment that attention mechanisms alone might miss.

2. Implicit Correspondence (Video Priors):

   • By treating the image group as a video sequence, the model inherits temporal coherence learned during pre-training, naturally handling semantic continuity.

3. Key Technical Innovations (RoPE Modules):
   To fuse explicit and implicit cues effectively, the authors introduce two specialized Rotary Positional Embedding (RoPE) variants:

   • Geometry-enhanced RoPE (Ge-RoPE):
     • Integrates displacement fields derived from VGGT.
     • Adjusts spatial positional indices based on geometric disparity and confidence levels.
     • Function: Aligns latent tokens with the geometric structure of the VGGT features, improving spatial accuracy in complex transformations.
   • Identity-enhanced RoPE (Identity-RoPE):
     • Uses a bounding-box-based matching strategy on segmentation masks.
     • Normalizes coordinates within object regions relative to the object's origin rather than the image frame.
     • Function: Ensures that corresponding object parts (e.g., a dog's left eye) share identical positional signatures across different images, preserving identity consistency.

3. Key Contributions

1. Framework: Proposes GroupEditing, the first training-based framework that treats related image groups as pseudo-video frames to achieve consistent multi-image editing.
2. Novel Modules: Introduces Ge-RoPE for fine-grained geometric alignment and Identity-RoPE for robust identity preservation across views.
3. Dataset & Benchmark:
   • GroupEditData: A large-scale dataset (>7K groups) with precise masks and detailed captions.
   • GroupEditBench: A dedicated benchmark with 800 image groups for evaluating group-level editing performance.
4. Applications: Demonstrates utility in 3D reconstruction (using consistent edits to improve point matching) and Image Customization (fine-tuning generative models on edited results).

4. Experimental Results

The authors evaluated GroupEditing against State-of-the-Art (SOTA) baselines (e.g., Anydoor, OminiControl, Edicho, StyleAligned) on the GroupEditBench.

• Quantitative Performance:
  • Local Editing: Outperformed baselines in CLIP-Score (0.3122), Aesthetic-Score (5.39), DINO-Score (0.8168), and Editing Consistency (0.9239).
  • Global Editing: Achieved the highest scores across all metrics, including a 0.9147 Editing Consistency score.
  • User Study: Ranked #1 by human evaluators in identity consistency, aesthetic quality, appearance fidelity, and overall quality.
• Ablation Studies:
  • Removing VGGT or Ge-RoPE significantly degraded spatial alignment and geometric accuracy.
  • Removing Identity-RoPE led to noticeable identity drift (inconsistency in object appearance) across the image group.
• Visual Quality: The method successfully handles complex scenarios like changing lighting, adding accessories, and style transfer while maintaining the subject's identity across different viewpoints.

5. Significance

This work represents a significant leap in generative image editing by moving beyond single-image constraints.

• Unified Editing: It solves the critical problem of maintaining coherence in multi-view or multi-pose scenarios, which is essential for digital avatars, e-commerce product visualization, and 3D reconstruction.
• Bridging Modalities: By effectively fusing explicit geometric matching (VGGT) with implicit video priors, it sets a new standard for how different types of visual correspondence can be integrated into diffusion models.
• Resource Availability: The release of GroupEditData and GroupEditBench fills a major gap in the community, providing the necessary resources for future research in consistent multi-image generation.