Research map / Image generation / Editing and style transfer
AdapEdit: Spatio-Temporal Guided Adaptive Editing Algorithm for Text-Based Continuity-Sensitive 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 AdapEdit, a training-free text-based image editing method that uses soft attention to adaptively set word-level temporal and pixel-level spatial guidance scales, enabling continuity-sensitive ("soft") edits on Stable Diffusion v1.4.
Problem
Existing text-guided diffusion editing methods handle discrete "hard" edits (add/remove/replace objects, styles) but struggle with continuity-sensitive "soft" edits (actions, poses, adjectives), and fixed guidance degrees cause artifacts and discontinuities, while fine-tuning-based methods cause lan
Approach
Replace uniform cross-attention guidance with a soft-attention strategy that computes per-word temporal truncation scales (flexible word-level temporal adjustment) and per-pixel spatial guidance weights (dynamic pixel-level spatial weighting, with a spatial interpolation weight lambda_S), all within a single forward diffusion pass without training.
Key results
- AdapEdit achieves the highest CLIP score (27.08) and CLIP directional similarity (0.13) among compared methods on the 700-sample InstructPix2Pix set.
- AdapEdit's LPIPS is 0.29, slightly worse than InstructPix2Pix (0.27) but far better than SDEdit (0.55) and MasaCtrl (0.41).
- Ablations show removing the FWT or DPS modules degrades editing of 'spread wings' and removing 'strawberries', per qualitative comparisons.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AdapEdit instead of Instruct P2P | SD-v1.4 image editing · InstructPix2Pix dataset (700 human-annotated samples) | CLIP score: 24.59 -> 27.08 | Helped |
| Use AdapEdit instead of Instruct P2P | SD-v1.4 image editing · InstructPix2Pix dataset (700 human-annotated samples) | CLIP directional: 0.04 -> 0.13 | Helped |
| Use AdapEdit instead of Instruct P2P | SD-v1.4 image editing · InstructPix2Pix dataset (700 human-annotated samples) | LPIPS: 0.27 -> 0.29 | Hurt |
| Add the flexible word-level temporal adjustment module (ablation) | AdapEdit · Continuity-sensitive posture editing (bird standing to spread wings) | Editing success | Helped |
| Add the dynamic pixel-level spatial weighting module (ablation) | AdapEdit · Removing strawberries from a plate | Editing success | Helped |
| Add pixel-level masking in DPS (ablation) | AdapEdit · Image editing with preservation of original characteristics | Consistency | Helped |
| Decrease the temporal hyperparameter λτ (ablation) | AdapEdit · Running horse editing | Semantic alignment | Helped |
| Vary λS[V] from 0.8 to 1.0 while keeping λS at 0.9 (ablation) | AdapEdit · Hamburger-to-cake editing | Editing amplitude: 0.8 to 1.0 | Mixed |
| Set λS to 0.98 and λS[V] to 1.0 (ablation) | AdapEdit · Hamburger-to-cake editing | Editing success: 0.98; 1.0 | Helped |
Limitations
- Quantitative evaluation is limited to a single 700-sample dataset and three metrics.
- Hyper-parameters (lambda_tau, lambda_S, lambda_S[V], alpha_m) must be manually chosen for different edits.
- LPIPS is slightly worse than InstructPix2Pix, indicating a trade-off between consistency and editing strength.
Experimental setup
Datasets: InstructPix2Pix evaluation dataset (700 human-annotated samples)
Benchmarks: InstructPix2Pix dataset (700 human-annotated samples)
Metrics: LPIPS CLIP score CLIP directional similarity
Baselines: Prompt-to-Prompt SDEdit InstructPix2Pix MasaCtrl
Builds on: Prompt-to-Prompt MasaCtrl Stable Diffusion ControlNet
Closest papers by meaning
- EditInfinity: Image Editing with Binary-Quantized Generative Models (NeurIPS 2025 · similarity 0.89)
- Textualize Visual Prompt for Image Editing via Diffusion Bridge (AAAI 2025 · similarity 0.88)
- Towards Efficient Diffusion-Based Image Editing with Instant Attention Masks (AAAI 2024 · similarity 0.88)
- TextCtrl: Diffusion-based Scene Text Editing with Prior Guidance Control (NeurIPS 2024 · similarity 0.88)
- BARET: Balanced Attention Based Real Image Editing Driven by Target-Text Inversion (AAAI 2024 · similarity 0.88)
- EditAR: Unified Conditional Generation with Autoregressive Models (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.