Research map / Image generation / Editing and style transfer
DiT4Edit: Diffusion Transformer for 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 DiT4Edit, a tuning-free diffusion-transformer image editing framework built on PixArt-α that combines DPM-Solver++ inversion, unified self-attention control, and patch merging for high-resolution and arbitrary-size editing.
Problem
Shape-aware object editing in high-resolution and arbitrary-size images is under-explored, since existing text-driven editing methods are UNet-based and rely on many DDIM inversion steps.
Approach
Adapts DPM-Solver++ inversion (fewer steps than DDIM inversion), a threshold-based unified attention control that switches between query features of source and target, and token-merging-style patch merging to a pretrained DiT (PixArt-α) for training-free editing.
Key results
- Quantitative table: DiT4Edit achieves FID 72.36 / 62.45 / 75.43 at 512x512 / 1024x1024 / 1024x2048, versus best UNet baseline FID around 84.73 at 512x512 (PnPInversion).
- Inference time of 5.15 s at 512x512, comparable to InfEdit (5.08 s) and much faster than PnPInversion (30.48 s) and MasaCtrl (19.76 s).
- Patch merging reduces 1024x1024 editing time from 41.52 s to 34.27 s and 1024x2048 from 122.41 s to 99.39 s, reported to keep quality comparable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiT4Edit instead of the strongest listed baseline | Image editing · 512×512, 1024×1024, and 1024×2048 | FID: 84.73 -> 72.36; 87.42 -> 62.45; 98.75 -> 75.43 | Helped |
| Use DiT4Edit instead of the strongest listed baseline | Image editing · 512×512, 1024×1024, and 1024×2048 | PSNR: 21.71 -> 22.85; 23.42 -> 29.75; 24.59 -> 27.46 | Helped |
| Use DiT4Edit instead of the strongest listed baseline | Image editing · 512×512, 1024×1024, and 1024×2048 | CLIP: 24.46 -> 25.39; 23.74 -> 26.97; 22.16 -> 25.66 | Helped |
| Use DiT4Edit instead of InfEdit | Image editing · 512×512 | Inference time: 5.08 -> 5.15 | Hurt |
| Add patches merging (ablation) | DiT4Edit · 512×512, 1024×1024, and 1024×2048 | Speed: 9.13 -> 6.01; 41.52 -> 34.27; 122.41 -> 99.39 | Helped |
| Add patches merging (ablation) | DiT4Edit · Image editing at the evaluated image sizes | Editing quality | No clear effect |
| Use DPM-Solver inversion instead of DDIM inversion (ablation) | DiT4Edit · Image editing, T=30 | Image editing quality | Helped |
Surprising findings
- Reported PSNR for MasaCtrl at 512x512 (17.35) is lower than several baselines, and the paper notes Pix2Pix-Zero PSNR varies strongly across sizes, indicating inconsistent baseline behavior.
Limitations
- T5 tokenizer occasionally fails at word segmentation, causing editing failures.
- Color inconsistencies compared to the original image.
- All baselines are UNet-based and no DiT-based editing baseline exists for comparison.
Experimental setup
Metrics: FID PSNR CLIP score inference time
Baselines: SDEdit InstructPix2Pix Pix2Pix-Zero MasaCtrl InfEdit PnPInversion
Builds on: PixArt-α DPM-Solver++ Prompt-to-Prompt MasaCtrl token merging
Closest papers by meaning
- Stable Flow: Vital Layers for Training-Free Image Editing (CVPR 2025 · similarity 0.89)
- Exploring Diffusion Transformer Designs via Grafting (NeurIPS 2025 · similarity 0.88)
- Enabling Instructional Image Editing with In-Context Generation in Large Scale Diffusion Transformer (NeurIPS 2025 · similarity 0.88)
- Dynamic Diffusion Transformer (ICLR 2025 · similarity 0.88)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.88)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.87)
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.