Research map / Image generation / Editing and style transfer
Object-Aware Inversion and Reassembly for 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 Object-aware Inversion and Reassembly (OIR), a training-free image editing method that searches an optimal DDIM inversion step per editing pair and reassembles separately edited regions, evaluated on new OIRBench datasets.
Problem
Text-driven diffusion image editing uses a fixed number of DDIM inversion steps for all edits, but optimal steps vary per editing pair, causing poor editing and concept mismatch in multi-object edits.
Approach
Search the optimal inversion step for each editing pair using a metric combining CLIP alignment of the edited region and preservation of the non-edited region, edit each pair separately with its own step, then reassemble the edited regions with the non-edited latent and apply re-inversion before final denoising.
Key results
- On the multi-object dataset, OIR achieves a CLIP score of 34.98 versus 32.97 for the best Stable Diffusion Inpainting variant (Method1) in the SDI comparison table.
- In a 58-participant user study on 15 multi-object images, OIR was preferred in 66.7% of selections.
- On single-object editing, the search metric reaches CLIP 27.49 and MS-SSIM 0.719, comparable to PNP (27.39, 0.680) and SDI (27.36, 0.706).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the OIR search metric instead of Plug-and-Play | Single-object editing · OIRBench single-object editing | CLIP score: 27.39 -> 27.49 | Helped |
| Use the OIR search metric instead of Stable Diffusion Inpainting | Single-object editing · OIRBench single-object editing | CLIP score: 27.36 -> 27.49 | Helped |
| Use the OIR search metric instead of DiffEdit | Single-object editing · OIRBench single-object editing | CLIP score: 21.23 -> 27.49 | Helped |
| Use OIR instead of Stable Diffusion Inpainting | Multi-object editing · OIRBench multi-object editing | CLIP score: 32.97 -> 34.98 | Helped |
| Use OIR instead of Stable Diffusion Inpainting | Multi-object editing · OIRBench multi-object editing | MS-SSIM: 0.516 -> 0.653 | Helped |
| Use OIR instead of Stable Diffusion Inpainting | Multi-object editing · OIRBench multi-object editing | LPIPS: 0.420 -> 0.329 | Helped |
| Run OIR candidate searches on multiple GPUs instead of one GPU | OIR · OIRBench multi-object editing | Average time cost: 298s -> 155s | Helped |
| Remove re-inversion during reassembly (ablation) | OIR · OIR ablation examples | Editing quality | Hurt |
| Set the reassembly step to a significantly large value, merging editing pairs early (ablation) | OIR · OIR ablation examples | Concept consistency | Hurt |
| Remove re-inversion and use an early large reassembly step (ablation) | OIR · OIR ablation examples | Editing quality | Hurt |
| Replace Grounded-SAM masks with attention-map masks (ablation) | OIR search metric · OIR mask-generator comparison | Mask accuracy | Hurt |
| Search for the Null-text Inversion cross-replace step instead of using its default (ablation) | Null-text Inversion with the search metric · Qualitative editing examples | Target editing alignment | Helped |
Surprising findings
- Simply selecting a good intermediate candidate from DDIM inversion steps often beats feature-injection editing methods.
- MS-SSIM favors methods that leave objects unedited, so it may not correlate with edit quality.
Limitations
- Requires additional inference time for the per-pair inversion step search.
- Effectiveness on video or other generic editing tasks is unverified.
- Relies on mask generator quality (Grounded-SAM by default), though attention-based masks are also usable.
Experimental setup
Datasets: Pexels, Unsplash, 500px images
Benchmarks: OIRBench multi-object OIRBench single-object
Metrics: CLIP score MS-SSIM LPIPS user preference
Baselines: DiffEdit Null-text Inversion Plug-and-Play Stable Diffusion Inpainting
Builds on: DDIM inversion Stable Diffusion Plug-and-Play Null-text Inversion Grounded-SAM
Closest papers by meaning
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.92)
- EditInfinity: Image Editing with Binary-Quantized Generative Models (NeurIPS 2025 · similarity 0.91)
- PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code (ICLR 2024 · similarity 0.91)
- DCI: Dual-Conditional Inversion for Boosting Diffusion-Based Image Editing (NeurIPS 2025 · similarity 0.90)
- PAIR Diffusion: A Comprehensive Multimodal Object-Level Image Editor (CVPR 2024 · similarity 0.90)
- Tuning-Free Inversion-Enhanced Control for Consistent Image Editing (AAAI 2024 · similarity 0.90)
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.