Research map / Image generation / Editing and style transfer
DNAEdit: Direct Noise Alignment for Text-Guided Rectified Flow 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 DNAEdit, a training-free rectified-flow image editing method that refines the Gaussian noise directly via velocity-gap alignment (DNA) plus a mobile velocity guidance (MVG), and introduces DNA-Bench, a long-prompt extension of PIE-Bench.
Problem
Re-noising based inversion in rectified flow (RF) text-guided image editing accumulates approximation drift, degrading reconstruction fidelity and editing quality; existing benchmarks also use short prompts unsuited to RF's long-text understanding.
Approach
Instead of inverting the clean latent step by step, DNA starts from random Gaussian noise and iteratively shifts it using the difference between the linear-path velocity and the model-predicted velocity, obtaining structured noise aligned with the source image; MVG then guides denoising with a moving reference trajectory computed from the source-target velocity difference.
Key results
- On PIE-Bench, DNAEdit (SD3.5 version) achieves the best average rank (2.50) with structure distance 14.19 and MSE 32.76, outperforming FTEdit (rank 3.50).
- The FLUX version of DNAEdit reaches PSNR 24.99 and structure distance 18.87 on PIE-Bench, versus 23.33 PSNR for FireFlow.
- On DNA-Bench, DNAEdit (SD3.5) attains the best average rank of 2.08 and highest CLIP whole/edited scores (28.90/23.66).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DNAEdit instead of FTEdit | SD3.5 image editing · PIE-Bench | CLIP Similarity Edited: 22.27 -> 22.71 | Helped |
| use DNAEdit instead of FTEdit | SD3.5 image editing · PIE-Bench | PSNR: 26.62 -> 26.66 | No clear effect |
| add DNA noise alignment (ablation) | interpolation-based inversion · PIE-Bench ablation | CLIP Similarity Whole: 25.73 -> 26.24 | Helped |
| add DNA noise alignment (ablation) | interpolation-based inversion · PIE-Bench ablation | CLIP Similarity Edited: 22.47 -> 23.08 | Helped |
| add residual offset during re-denoising (ablation) | interpolation-based method · PIE-Bench ablation | Structure Distance: 31.90 -> 24.93 | Helped |
| add residual offset during re-denoising (ablation) | interpolation-based method · PIE-Bench ablation | PSNR: 22.61 -> 23.75 | Helped |
| add MVG (ablation) | DNA + ResOffset · PIE-Bench ablation | Structure Distance: 33.98 -> 18.87 | Helped |
| add MVG (ablation) | DNA + ResOffset · PIE-Bench ablation | PSNR: 21.97 -> 24.99 | Helped |
| add MVG (ablation) | DNA + ResOffset · PIE-Bench ablation | CLIP Similarity Whole: 26.40 -> 25.79 (-0.61) | Hurt |
| add MVG (ablation) | DNA + ResOffset · PIE-Bench ablation | CLIP Similarity Edited: 23.23 -> 22.87 | Hurt |
| add DNA and ResOffset to interpolation (ablation) | interpolation-based method · PIE-Bench ablation | Structure Distance: 31.90 -> 33.98 | Hurt |
Surprising findings
- Inversion-based methods achieve better background preservation with long prompts than with short prompts (e.g., FireFlow PSNR 23.33 dB on PIE-Bench vs 23.57 dB on DNA-Bench).
Limitations
- May fail on editing cases outside the foundation T2I model's prior.
- Relies on long-prompt descriptions; benchmark construction depends on GPT-4o generated captions.
- MVG coefficient eta must be tuned, since excessive values can cause editing failures.
Experimental setup
Datasets: PIE-Bench
Benchmarks: PIE-Bench DNA-Bench
Metrics: structure distance PSNR LPIPS MSE SSIM CLIP similarity average rank
Baselines: PnP MasaCtrl InfEdit InsP2P RF-Inversion RFEdit FireFlow FlowEdit FTEdit
Builds on: rectified flow FlowEdit RF-Inversion FireFlow RF-Solver
Closest papers by meaning
- Schedule Your Edit: A Simple yet Effective Diffusion Noise Schedule for Image Editing (NeurIPS 2024 · similarity 0.87)
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.87)
- Noise Map Guidance: Inversion with Spatial Context for Real Image Editing (ICLR 2024 · similarity 0.86)
- Editable Noise Map Inversion: Encoding Target-image into Noise For High-Fidelity Image Manipulation (ICML 2025 · similarity 0.86)
- Why Do DiT Editors Drift? Plug-and-Play Low Frequency Alignment in VAE Latent Space (NeurIPS 2026 · similarity 0.86)
- SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing (NeurIPS 2025 · 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.