Research map / Image generation / Editing and style transfer
FeedEdit: Text-Based Image Editing with Dynamic Feedback Regulation
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 FeedEdit, a training-free text-based image editing method that feedback-regulates per-step, per-phrase text guidance intensity using frequency-filtered feature differences and a PI controller.
Problem
Text-based image editing with diffusion models uses a fixed text guidance (editing) intensity across all denoising steps, which causes over-editing and under-editing and performs poorly on multi-function edits.
Approach
Perceive the editing degree at each step from frequency-filtered value and attention feature differences between editing and reconstruction branches, then use a non-linear proportional-integral controller to produce step-specific control signals applied to cross-attention keys and values per part-of-speech phrase group.
Key results
- Single-function editing: CLIP-T 0.3024 vs 0.2770 for best baseline SDPInV, and FID 65.02 vs 77.16 for InfEdit.
- Multi-function editing: CLIP-T 0.2607 vs 0.2367 for InfEdit, and FID 81.70 vs 98.15 for InfEdit.
- Ablation: removing the phrase-level regulating strategy lowers CLIP-T by about 7.4%, and removing frequency-domain filtering lowers CLIP-T to 0.2588 from 0.2857.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FeedEdit instead of SDPInV | Stable Diffusion 1.5 image editing · Single-function editing | CLIP-T: 0.2770 -> 0.3024 | Helped |
| Use FeedEdit instead of SDPInV | Stable Diffusion 1.5 image editing · Single-function editing | LPIPS: 0.1525 -> 0.1311 | Helped |
| Use FeedEdit instead of InfEdit | Stable Diffusion 1.5 image editing · Multi-function editing | CLIP-T: 0.2367 -> 0.2607 | Helped |
| Use FeedEdit instead of InfEdit | Stable Diffusion 1.5 image editing · Multi-function editing | FID: 98.15 -> 81.70 | Helped |
| Add frequency-domain filtering for editing-degree perception (ablation) | FeedEdit · Image editing | CLIP-T: 0.2588 -> 0.2857 | Helped |
| Use L1-norm instead of L2-norm for editing-degree perception (ablation) | FeedEdit · Image editing | CLIP-T: 0.2776 -> 0.2857 | Helped |
| Use a non-linear PI controller instead of linear mapping (ablation) | FeedEdit · Image editing | CLIP-T: 0.2514 -> 0.2857 | Helped |
| Add the proportional term (ablation) | FeedEdit PI controller · Image editing | CLIP-T: 0.2614 -> 0.2857 | Helped |
| Add the integral term (ablation) | FeedEdit PI controller · Image editing | CLIP-T: 0.2780 -> 0.2857 | Helped |
| Add phrase-level regulating (ablation) | FeedEdit · Image editing | CLIP-T: 0.2647 -> 0.2857 | Helped |
Surprising findings
- The naive strategy of scaling global text guidance improves single-function editing but performs poorly on multi-function editing, motivating phrase-level regulation.
Limitations
- Ideal editing degree thresholds are set from statistics on specific datasets and require manual tuning for multi-function editing.
- Evaluated on only 100 test samples with CLIP-based and FID metrics, with no human evaluation.
Experimental setup
Datasets: InstructPix2Pix dataset MagicBrush self-collected samples
Benchmarks: self-collected 100-sample test set (60 single-function, 40 multi-function)
Metrics: CLIP-T CLIP-I LPIPS FID ImageReward
Baselines: Imagic MasaCtrl InstructPix2Pix DAC InfEdit SDPInV
Builds on: Prompt-to-Prompt MasaCtrl Classifier-free guidance DDIM
Closest papers by meaning
- AutoEdit: Automatic Hyperparameter Tuning for Image Editing (NeurIPS 2025 · similarity 0.88)
- Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · similarity 0.88)
- AdapEdit: Spatio-Temporal Guided Adaptive Editing Algorithm for Text-Based Continuity-Sensitive Image Editing (AAAI 2024 · similarity 0.87)
- EditInfinity: Image Editing with Binary-Quantized Generative Models (NeurIPS 2025 · similarity 0.87)
- ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.87)
- Enabling Instructional Image Editing with In-Context Generation in Large Scale Diffusion Transformer (NeurIPS 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.