Research map / Image generation / Editing and style transfer
MoEdit: On Learning Quantity Perception for Multi-object 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 MoEdit, an auxiliary-free multi-object image editing framework on SDXL that uses Feature Compensation and Quantity Attention modules to preserve object counts and attributes during style transfer, object reinvention and background regeneration.
Problem
Multi-object image editing with diffusion models often yields inconsistent quantity perception and object attribute aliasing, and existing fixes rely on auxiliary tools such as masks, LLMs or bounding boxes.
Approach
A text-guided Feature Compensation module enhances CLIP image features to separate per-object attributes, and a Quantity Attention module injected into U-Net block B4 provides global quantity control, all without masks or LLMs.
Key results
- MoEdit achieves 75.44% numerical accuracy in the subjective evaluation versus 11.35% for TurboEdit and 3.87% for Emu2 (Emu2 scores slightly higher on numerical accuracy in the table's breakdown; MoEdit leads MOS at 70.34 vs 59.33 for TurboEdit).
- MoEdit obtains the best HyperIQA (75.66), CLIP Score (0.3254), Q-Align (4.8763) and AesBench (4.6172) among compared methods, with best NIQE (2.749) too.
- Ablation: removing FeCom drops numerical accuracy to 25.67 from 86.79 with full model; removing QTTN drops it to 82.31.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MoEdit instead of TurboEdit | SDXL multi-object image editing · Multi-object image editing evaluation | HyperIQA: 71.72 -> 75.66 | Helped |
| use MoEdit instead of TurboEdit | SDXL multi-object image editing · Multi-object image editing evaluation | LPIPS: 0.2427 -> 0.2557 | Hurt |
| use MoEdit instead of Emu2 | SDXL multi-object image editing · Multi-object image editing evaluation | Numerical accuracy: 92.82 -> 75.44 | Hurt |
| add the QTTN module (ablation) | MoEdit · Multi-object image editing evaluation | Numerical accuracy: 25.67 -> 86.79 | Helped |
| add the FeCom module (ablation) | MoEdit · Multi-object image editing evaluation | CLIP Score: 0.3052 -> 0.3274 | Helped |
| add the FeCom module (ablation) | MoEdit · Multi-object image editing evaluation | LPIPS: 0.2777 -> 0.2731 | Hurt |
| insert QTTN into U-Net block B4 rather than other tested single blocks (ablation) | MoEdit · Multi-object image editing evaluation | quantity perception and editability | Helped |
| decrease λ from 1 toward 0 (ablation) | MoEdit FeCom module · Multi-object image editing evaluation | preservation of edited visual attributes | Hurt |
Surprising findings
- Emu2 slightly outperforms MoEdit on the subjective numerical accuracy metric, which the authors attribute to Emu2's clustering strategy.
Limitations
- Robustness in complex interaction scenarios is not addressed, as acknowledged by the authors.
- Small evaluation set (30 test images) and in-house dataset that is not yet released.
- Emu2 beats it on subjective numerical accuracy.
Experimental setup
Datasets: in-house multi-object image dataset (over 1,200 images)
Benchmarks: in-house multi-object editing test set (30 images, 50-150 instructions each)
Metrics: NIQE HyperIQA CLIP Score LPIPS Q-Align AesBench MOS Numerical Accuracy
Baselines: SSR-Encoder lambda-Eclipse IP-Adapter Blip-diffusion MS-diffusion Emu2 TurboEdit
Builds on: Stable Diffusion XL CLIP Null-text inversion
Closest papers by meaning
- PAIR Diffusion: A Comprehensive Multimodal Object-Level Image Editor (CVPR 2024 · similarity 0.87)
- Object-Aware Inversion and Reassembly for Image Editing (ICLR 2024 · similarity 0.86)
- Multi-Reward as Condition for Instruction-based Image Editing (ICLR 2025 · similarity 0.85)
- An Item Is Worth a Prompt: Versatile Image Editing with Disentangled Control (AAAI 2025 · similarity 0.85)
- PS-Diffusion: Photorealistic Subject-Driven Image Editing with Disentangled Control and Attention (CVPR 2025 · similarity 0.84)
- OmniEdit: Building Image Editing Generalist Models Through Specialist Supervision (ICLR 2025 · similarity 0.84)
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.