Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
DNAEdit: Direct Noise Alignment for Text-Guided Rectified Flow Editing
Authors: Chenxi Xie, Minghan Li, Shuai Li, Yuhui Wu, Qiaosi Yi, Lei Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our DNAEdit achieves superior performance to state-of-the-art text-guided editing methods. Experiments on PIE-Bench and DNA-Bench show that our DNAEdit method strikes a better balance between fidelity and editability, demonstrating the best performance. |
| Researcher Affiliation | Collaboration | 1The Hong Kong Polytechnic University, 2OPPO Research Institute, 3Harvard University |
| Pseudocode | Yes | Algorithm 1 Direct Noise Alignment (DNA) Input: Number of optimization steps T, Source image ZT , Source text ψsrc, Timesteps {σt}0 t=T , RF model vθ with parameters θ, Randomly sampled start noise ST N(0, 1). Output: Noisy latents {Zt}0 t=T 1 and residual offset { x DNA t }0 t=T 1, Terminal noise sample S0 |
| Open Source Code | No | Our code, model, and benchmark will be made publicly available. We will release the codes and new data if the paper is accepted. |
| Open Datasets | Yes | Experiments on PIE-Bench and DNA-Bench show that our DNAEdit method strikes a better balance between fidelity and editability, demonstrating the best performance. PIE-Bench [6] has an average prompt length of 9.46 words. |
| Dataset Splits | No | The PIE-bench consists of 700 natural and artificial images to evaluate editing methods across 9 distinct dimensions. It provides the source and target prompts for each image, along with the editing area masks to assess background preservation and local editing ability. |
| Hardware Specification | No | Justification: We detail the type of compute resources in appendix. |
| Software Dependencies | No | Two versions of DNAEdit are provided, which are based on FLUX-dev [9] and SD3.5-medium [19], respectively. |
| Experiment Setup | Yes | In both versions, the MVG coefficient η is fixed at 0.8. Detailed hyper-parameter settings can be found in the Appendix. |