Research map / Image generation / Editing and style transfer
Unveil Inversion and Invariance in Flow Transformer for Versatile 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 a tuning-free image editing framework for flow transformers (SD3.5 MM-DiT) with a two-stage Euler rectified-flow inversion (fixed-point velocity refinement plus per-step velocity compensation) and AdaLN text-feature replacement for invariance control.
Problem
Tuning-free image editing with flow-based transformer text-to-image models is hampered by poor Euler inversion (worse than DDIM inversion in diffusion) and by invariance-control methods built for U-Net attention that cannot reconcile rigid and non-rigid edits.
Approach
Analyze Euler inversion as a DDIM-like first-order solver that suffers more from velocity approximation error, and fix it with fixed-point iteration plus per-step velocity compensation; for invariance, replace unedited-token text features inside MM-DiT's AdaLN modulation (before timestep S) rather than injecting self-attention maps.
Key results
- On PIE benchmark, the method reaches PSNR 26.62, SSIM 91.50 and CLIP Whole 25.74 with structure distance 18.17, versus RFinv structure distance 32.62 and PnP 24.29.
- Fixed-point iteration ablation: structure distance drops from 48.49 (plain Euler, Iter 0) to 14.92 (Iter 3), with PSNR rising from 20.78 to 27.51.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add fixed-point iteration to Euler inversion (0 to 1 iteration) (ablation) | Two-stage flow inversion · PIE benchmark | PSNR: 20.78 -> 26.98 | Helped |
| Add fixed-point iteration to Euler inversion (0 to 1 iteration) (ablation) | Two-stage flow inversion · PIE benchmark | CLIP Similarity (Edited): 21.39 -> 21.47 | Helped |
| Increase fixed-point iterations from 1 to 3 (ablation) | Two-stage flow inversion · PIE benchmark | CLIP Similarity (Edited): 21.47 -> 21.73 | Helped |
| Increase fixed-point iterations from 1 to 3 (ablation) | Two-stage flow inversion · PIE benchmark | PSNR: 26.98 -> 27.51 | Helped |
| Replace unchanged text features in AdaLN while keeping edited text features intact | MM-DiT image editing · Versatile rigid and non-rigid editing | Editing flexibility and preservation of non-target contents | Helped |
| Use self-attention injection for more than 20% of timesteps | MM-DiT image editing · Non-rigid editing, including sitting to standing | Non-rigid editing ability | Hurt |
Surprising findings
- Plain Euler inversion on SD3.5 performs worse than DDIM inversion on SD1.5 despite the similar formulation, as shown in Figure 3.
Limitations
- Invariance control relies on the inversion matching the generation process; out-of-domain real images yield poor inversion and editing.
- Fixed-point iteration increases computation by requiring more transformer inferences.
Experimental setup
Datasets: PIE benchmark
Benchmarks: PIE benchmark
Metrics: PSNR LPIPS MSE SSIM CLIP similarity (whole) CLIP similarity (edited) structure distance
Baselines: P2P PnP MasaCtrl InsP2P InfEdit RFinv
Builds on: DDIM inversion rectified flow Prompt-to-Prompt Plug-and-Play Direct Inversion
Closest papers by meaning
- SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing (NeurIPS 2025 · similarity 0.90)
- Stable Flow: Vital Layers for Training-Free Image Editing (CVPR 2025 · similarity 0.90)
- Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations (ICLR 2025 · similarity 0.89)
- Taming Rectified Flow for Inversion and Editing (ICML 2025 · similarity 0.89)
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.89)
- EditInfinity: Image Editing with Binary-Quantized Generative Models (NeurIPS 2025 · similarity 0.88)
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.