Research map / Image generation / Editing and style transfer
ParallelEdits: Efficient Multi-Aspect Text-Driven Image Editing with Attention Grouping
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 ParallelEdits, an inversion-free multi-branch diffusion editor that groups attention-mapped aspects into parallel branches (rigid, non-rigid, global) built on LCM, and introduces the PIE-Bench++ multi-aspect editing benchmark.
Problem
Text-driven image editing methods handle a single aspect well but degrade when multiple objects or attributes must be edited simultaneously; sequential application is slow and accumulates errors.
Approach
Group editable aspects by attention-map overlap and edit type into parallel target branches that are calibrated against each other within one diffusion trajectory, using cross-branch attention injection (P2P-style replace/refine for rigid edits, K/V retention for non-rigid edits), so the cost does not scale with the number of aspects.
Key results
- On PIE-Bench++, ParallelEdits reaches D-CLIP 20.70% and AspAcc-LLaVA 65.19%, the best among compared methods in Table 1.
- ParallelEdits runs in 4.98 s/sample versus 121-123 s for sequential InfEdit and PnP variants, while being slightly slower than single-pass InfEdit (2.22 s for DI* etc.; InfEdit 29.70 s in table).
- Ablation shows removing grouping and auxiliary branch drops AspAcc-CLIP from 51.05% to 48.08% and LLaVA accuracy from 65.19% to 61.22%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ParallelEdits instead of the strongest reported baseline | PIE-Bench++ image editing · PIE-Bench++ | AspAcc-CLIP (%): 48.20 -> 51.05 | Helped |
| use ParallelEdits instead of the strongest reported baseline | PIE-Bench++ image editing · PIE-Bench++ | AspAcc-LLaVA (%): 63.80 -> 65.19 | Helped |
| use ParallelEdits instead of the strongest reported baseline | PIE-Bench++ image editing · PIE-Bench++ | D-CLIP (%): 13.23 -> 20.70 | Helped |
| use ParallelEdits instead of InfEdit | PIE-Bench++ image editing · PIE-Bench++ | Eff. (secs/sample): 2.22 -> 4.98 | Hurt |
| add aspect categorization, aspect grouping, and auxiliary branches (ablation) | ParallelEdits · PIE-Bench++ | D-CLIP: 10.45 -> 20.70 | Helped |
| add aspect categorization (ablation) | ParallelEdits without aspect categorization · PIE-Bench++ | D-CLIP: 11.97 -> 20.70 | Helped |
| add aspect grouping and auxiliary branches (ablation) | ParallelEdits without aspect grouping and auxiliary branches · PIE-Bench++ | D-CLIP: 12.33 -> 20.70 | Helped |
Surprising findings
- Sequential application of single-aspect editors can undo earlier edits and accumulate artifacts, and sometimes performs worse than the vanilla single-pass method on aspect accuracy.
- Multimodal LLM GPT-4V fails to preserve source image content when used for editing.
Limitations
- Cannot edit text rendered within images (failure case shown).
- Fails on dramatic background changes.
- Relies on LLaVA and CLIP judges for aspect accuracy metrics, which can misjudge (e.g., color or background matches).
Experimental setup
Datasets: PIE-Bench
Benchmarks: PIE-Bench++ PIE-Bench
Metrics: CLIP score D-CLIP AspAcc-CLIP AspAcc-LLaVA PSNR LPIPS MSE SSIM runtime (secs/sample)
Baselines: DDIM+MasaCtrl DDIM+Prompt-to-Prompt DDIM+Plug-and-Play StyleDiffusion+P2P Null-text Inversion+P2P DirectInversion+PnP InfEdit Rich-text GPT-4V
Builds on: Prompt-to-Prompt Direct Inversion InfEdit MasaCtrl Latent Consistency Model
Closest papers by meaning
- ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.89)
- LEDITS++: Limitless Image Editing using Text-to-Image Models (CVPR 2024 · similarity 0.87)
- InsightEdit: Towards Better Instruction Following for Image Editing (CVPR 2025 · similarity 0.87)
- UltraEdit: Instruction-based Fine-Grained Image Editing at Scale (NeurIPS 2024 · similarity 0.87)
- An Item Is Worth a Prompt: Versatile Image Editing with Disentangled Control (AAAI 2025 · similarity 0.86)
- TextCtrl: Diffusion-based Scene Text Editing with Prior Guidance Control (NeurIPS 2024 · similarity 0.86)
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.