notesum.ai
Published at December 5SwiftEdit: Lightning Fast Text-Guided Image Editing via One-Step Diffusion
cs.CV
Released Date: December 5, 2024
Authors: Trong-Tung Nguyen1, Quang Nguyen, Khoi Nguyen, Anh Tran, Cuong Pham
Aff.: 1VinAI Research
![[Uncaptioned image]](https://arxiv.org/html/2412.04301v1/x1.png)
| Type | Method | Background Preservation | CLIP Semantics | Runtime | ||
|---|---|---|---|---|---|---|
| PSNR | MSE | Whole | Edited | (seconds) | ||
| Multi-step (50 steps) | DDIM + P2P | 17.87 | 219.88 | 25.01 | 22.44 | 25.98 |
| NT-Inv + P2P | 27.03 | 35.86 | 24.75 | 21.86 | 134.06 | |
| DDIM + MasaCtrl | 22.17 | 86.97 | 23.96 | 21.16 | 23.21 | |
| Direct Inversion + MasaCtrl | 22.64 | 81.09 | 24.38 | 21.35 | 29.68 | |
| DDIM + P2P-Zero | 20.44 | 144.12 | 22.80 | 20.54 | 35.57 | |
| Direct Inversion + P2P-Zero | 21.53 | 127.32 | 23.31 | 21.05 | 35.34 | |
| DDIM + PnP | 22.28 | 83.64 | 25.41 | 22.55 | 12.62 | |
| Direct Inversion + PnP | 22.46 | 80.45 | 25.41 | 22.62 | 12.79 | |
| Few-steps (4 steps) | ReNoise (SDXL Turbo) | 20.28 | 54.08 | 24.29 | 21.07 | 5.11 |
| TurboEdit | 22.43 | 9.48 | 25.49 | 21.82 | 1.32 | |
| ICD (SD 1.5) | 26.93 | 3.32 | 22.42 | 19.07 | 1.62 | |
| One-step | SwiftEdit (Ours) | 23.33 | 6.60 | 25.16 | 21.25 | 0.23 |
| SwiftEdit (Ours with GT masks) | 23.31 | 6.18 | 25.56 | 21.91 | 0.23 | |