Research map / Image generation / Editing and style transfer
Prompt Tuning Inversion for Text-driven Image Editing Using Diffusion Models
Wenkai Dong, Song Xue, Xiaoyue Duan, Shumin Han
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace DDIM inversion with Prompt Tuning Inversion for reconstruction | Stable Diffusion editing · COCO val reconstruction (w=7.5) | PSNR: 13.64 -> 25.71 | Improved |
| optimize conditional embedding (PTI) instead of unconditional embedding (NTI) | inversion · COCO val reconstruction | PSNR: 24.45 -> 25.71 | Improved |
| optimize conditional rather than unconditional embedding at larger learning rate and iterations (ablation) | inversion · COCO val, lr 0.1, 5 iterations | PSNR: 25.91 -> 25.97 | No clear effect |
| use more optimization iterations N from 1 to 5 (ablation) | PTI, lr 0.01 · COCO val reconstruction | PSNR: 19.36 -> 25.71 | Improved |
| increase guidance scale in DDIM forward and sampling together (ablation) | DDIM inversion · reconstruction PSNR | PSNR: 21.94 (w=1.0) -> 12.12 (w=5.0) | Hurt |
| use matching guidance scale between DDIM forward and sampling (ablation) | DDIM inversion · reconstruction PSNR | PSNR: 13.64 -> 21.36 (w_enc 0.0: w_dec 7.5 vs 0.0) | Improved |
| interpolate target and optimized condition embedding with eta 0.9 instead of eta 1.0 (ablation) | PTI editing · ImageNet CSFID-LPIPS trade-off | CSFID-LPIPS trade-off | Improved |
| reduce interpolation ratio eta below 0.9 (ablation) | PTI editing · ImageNet CSFID-LPIPS trade-off | CSFID-LPIPS trade-off | Hurt |
| lower PTI learning rate from 0.1 to 0.05 or 0.01 (ablation) | PTI editing · ImageNet CSFID-LPIPS trade-off | CSFID-LPIPS trade-off | Hurt |
| use latent interpolation instead of condition interpolation (ablation) | PTI editing · qualitative | image quality | Hurt |
| use cross-attention-map masks (DDIM-Edit) instead of DiffEdit-generated masks | DDIM-Edit · ImageNet CSFID-LPIPS trade-off | CSFID-LPIPS trade-off | Improved |
| add Prompt Tuning Inversion on top of DDIM-Edit | DDIM-Edit · ImageNet CSFID-LPIPS trade-off | CSFID-LPIPS trade-off | Improved |
Closest papers by meaning
- EditInfinity: Image Editing with Binary-Quantized Generative Models (NeurIPS 2025 · similarity 0.93)
- Object-Aware Inversion and Reassembly for Image Editing (ICLR 2024 · similarity 0.92)
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.92)
- Textualize Visual Prompt for Image Editing via Diffusion Bridge (AAAI 2025 · similarity 0.91)
- PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code (ICLR 2024 · similarity 0.90)
- Prompting Hard or Hardly Prompting: Prompt Inversion for Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.90)
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.
