Research map / Image generation / Text-to-image diffusion
ReFlex: Text-Guided Editing of Real Images in Rectified Flow via Mid-Step Feature Extraction and Attention Adaptation
Jimyeong Kim, Jungwon Park, Yeji Song, Nojun Kwak, Wonjong Rhee
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use mid-step feature extraction instead of extracting at every early step of the inverted latent (ablation) | ReFlex · PIE-Bench / Wild-TI2I-Real | source IoU | Improved |
| add I2T-CA adaptation (ablation) | ReFlex · PIE-Bench / Wild-TI2I-Real | text alignment | Improved |
| add I2I-SA adaptation (ablation) | ReFlex · PIE-Bench / Wild-TI2I-Real | text alignment | Improved |
| extract features at t'=14 instead of fully inverted latent t'=28 (ablation) | ReFlex · qualitative | source structure: t'=28 -> t'=14 | Improved |
| extract features at t'=14 instead of t'=4 (ablation) | ReFlex · qualitative | image quality: t'=4 -> t'=14 | Improved |
| increase k in I2I-SA adaptation (ablation) | ReFlex · qualitative | editability vs structure | Mixed |
| increase alpha from 1 in I2T-CA adaptation (ablation) | ReFlex · qualitative | text alignment | Mixed |
| apply latent blending for T steps instead of 0.7T (ablation) | ReFlex (k=20) · PIE-Bench | Background PSNR: 24.21 -> 28.05 | Improved |
| apply latent blending for T steps instead of 0.7T (ablation) | ReFlex (k=20) · PIE-Bench | CLIP similarity (whole): 26.51 -> 26.30 | Hurt |
| increase k from 20 to 40 (ablation) | ReFlex (m=0.7T) · PIE-Bench | CLIP similarity (edited region): 23.17 -> 23.37 | Improved |
| increase k from 20 to 40 (ablation) | ReFlex (m=0.7T) · PIE-Bench | Background PSNR: 24.21 -> 24.03 | Hurt |
| use ReFlex on FLUX instead of RF-Edit | FLUX editing · PIE-Bench | CLIP similarity (whole): 25.41 -> 26.51 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- FireFlow — FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing (ICML 2025) · better on 2 of 3 shared results
- FlowEdit — FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models (ICCV 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- BARET: Balanced Attention Based Real Image Editing Driven by Target-Text Inversion (AAAI 2024 · similarity 0.89)
- Stable Flow: Vital Layers for Training-Free Image Editing (CVPR 2025 · similarity 0.89)
- EditInfinity: Image Editing with Binary-Quantized Generative Models (NeurIPS 2025 · similarity 0.88)
- Textualize Visual Prompt for Image Editing via Diffusion Bridge (AAAI 2025 · similarity 0.88)
- LEDITS++: Limitless Image Editing using Text-to-Image Models (CVPR 2024 · similarity 0.88)
- SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing (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.
