1. Reduces Inversion Error Accumulation: Unlike conventional diffusion/RF inversion (which approximates later noisy latents with current ones), DNA optimizes Gaussian noise directly in the noise domain. It adjusts noise via velocity field differences, leveraging RF’s noise-clean image interpolation property to fundamentally reduce errors.

2. Balances Editing & Background Preservation: The supporting MVG module controls text-guided generation, ensuring target editability while maximizing original background retention.

3. Introduces DNA-Bench: A long-prompt benchmark addressing simple-prompt limitations in existing tests, enabling comprehensive evaluation of advanced editing models.

Edting samples are from Five-Bench

• Add gradio demo.
• Add Video Editing on Wan 2.1

1. Environment: torch=2.3.1 diffusers==0.30.1
2. Download the checkpoint SD-3.5-meidum or FLUX.1 Dev on Huggingface

1. Modify the path of PIE-Bench and your Diffusion models checkpoints in scripts/run_script_dnaedit.py

2. run: running on SD3.5-medium

   python scripts/run_script_dnaedit.py --device_number 0 --exp_yaml configs/DNAEdit_SD3_exp.yaml --save ./output
   

   or running on FLUX.1 Dev

   python scripts/run_script_dnaedit.py --device_number 0 --exp_yaml configs/DNAEdit_FLUX_exp.yaml --save ./output
   

3. If you want to inference on DNA-Bench, you can simply modify the json path as 'DNA-Bench/long_mapping_file.json' .

1. Configure Model Path

   Edit app.py and set your FLUX model path:

   pipe = FluxPipeline.from_pretrained(
       "path/to/your/FLUX-dev",  # Change this
       torch_dtype=torch.float16
   )

2. Launch gradio Demo

   python app.py

   Open http://localhost:7860 in your browser.

3. How to use

   Step 1 : Load Model and Upload Image

   1. First, load the model by selecting your preferred model and clicking the ⬇️ Load Model button
   2. Wait until the status shows ✅ Model loaded successfully
   3. Upload the image you want to edit

   Step 2: Set Prompts

   Provide a description of your input image as the Source Prompt.

   ⚠️ Important Note: The more accurate your image description, the more precise the editing will be and the higher the fidelity to the original image.

   You can use existing VLM to generate accurate captions such as Qwen-VL, Doubao.

   You also need to provide a description of your desired edited result as the Target Prompt.

   💡 Tip: Keep the grammatical structure as similar as possible to the Source Prompt, only changing the specific elements you want to edit.

   Step 3 : Configure Parameters and Start Editing

   Set the following 4 parameters before editing:

   Parameter Setting Tips

   T_step (Total Steps)

   • For FLUX models: Set to 28
   • For other models: Adjust according to model requirements

   T_start (Starting Skip Steps)

   • Recommended: 4
   • Effect: Balances editability and faithfulness
   • Lower values (e.g., 2-3): Stronger editing capability, more significant changes
   • Higher values (e.g., 5-8): More faithful to original image, subtle edits

   CFG

   • Recommended: 2.0 - 2.5
   • Warning: Values too small (<1.5) or too large (>4.0) will produce poor quality results
   • Optimal range provides best balance between prompt adherence and image quality

   MVG Coefficient

   • Recommended: ~0.85 (0.80 - 0.90)
   • Effect: Controls editing strength
   • Higher values (0.9-1.0): Stronger editing effects, more dramatic changes
   • Lower values (0.7-0.8): More faithful to original image, subtle modifications

   Step 4 : Start Editing

   1. Click the 🚀 Start Editing button
   2. Monitor the progress bar during processing
   3. View the edited result in the Editing Result panel

@article{xie2025dnaedit,
  title={DNAEdit: Direct Noise Alignment for Text-Guided Rectified Flow Editing},
  author={Xie, Chenxi and Li, Minghan and Li, Shuai and Wu, Yuhui and Yi, Qiaosi and Zhang, Lei},
  journal={arXiv preprint arXiv:2506.01430},
  year={2025}
}

This project is released under the Apache 2.0 license.

The code is largely based on FlowEdit whom we thank for their excellent work.

If you have any questions, please contact: chenxi.xie@connect.polyu.hk