Research map / Detection and segmentation / Open-set and few-shot detection
Style-Editor: Text-driven Object-centric Style 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 Style-Editor, a text-only object-centric style editing network trained per image with CLIP-based patch-wise co-directional loss, text-matched patch selection, and adaptive background preservation loss.
Problem
Text-driven style editing of specific objects in an image without reference style images or segmentation masks, while preserving object structure and background fidelity.
Approach
Uses CLIP to locate object patches from the source text (TMPS plus a pre-fixed foreground region, PRS), applies a patch-wise directional loss with a Jensen-Shannon patch distribution consistency term (PCD), and an adaptive MS-SSIM+L1 background preservation loss on dynamically identified background regions.
Key results
- Style-Editor reaches SimF 0.33 and ConF 3.75 on foreground, beating Text2LIVE (0.32, 4.13).
- Background PSNR 27.65 and SSIM 0.90 versus Text2LIVE's 24.69 and 0.87.
- Ablation shows removing the ABP loss sharply raises background L1 (0.48 vs 0.10).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Style-Editor instead of Text2LIVE | text-driven object-centric style editing · MSCOCO 2017 (160 stylized images) | SimF: 0.32 -> 0.33 | Helped |
| use Style-Editor instead of Text2LIVE | text-driven object-centric style editing · MSCOCO 2017 (160 stylized images) | L1B: 0.14 -> 0.10 | Helped |
| add the Adaptive Background Preservation loss (ABP) (ablation) | Style-Editor · MSCOCO 2017 ablation | L1B: 0.48 -> 0.10 | Helped |
| add the patch distribution consistency loss (L_con) (ablation) | Style-Editor · MSCOCO 2017 ablation | SimF: 0.32 -> 0.33 | Helped |
| add the patch distribution consistency loss (L_con) (ablation) | Style-Editor · MSCOCO 2017 ablation | ConF: 4.16 -> 3.75 | Helped |
Limitations
- Relies on CLIP patch-level zero-shot localization, which may fail for ambiguous source texts.
- Per-image optimization of about 45 seconds per prompt limits throughput.
- Background changes remain noticeable when only the directional loss is used (ablation).
Experimental setup
Datasets: MSCOCO 2017
Benchmarks: MSCOCO 2017 subset (16 images x 10 style texts, 160 stylized images)
Metrics: SimF ConF L1B ConB StyB SSIMB DISTSB PSNRB
Baselines: FlexIT ZeCon Null-text inversion Instruct Pix2Pix Plug and Play pix2pix-zero LEDITS++ local-prompt-mixing StylerDALLE CLIPstyler Text2LIVE Blended Diffusion
Builds on: StyleGAN-NADA directional loss CLIPstyler Neural Style Transfer manifold augmentation
Closest papers by meaning
- StyleStudio: Text-Driven Style Transfer with Selective Control of Style Elements (CVPR 2025 · similarity 0.86)
- TextCtrl: Diffusion-based Scene Text Editing with Prior Guidance Control (NeurIPS 2024 · similarity 0.85)
- Object-Aware Inversion and Reassembly for Image Editing (ICLR 2024 · similarity 0.85)
- AdapEdit: Spatio-Temporal Guided Adaptive Editing Algorithm for Text-Based Continuity-Sensitive Image Editing (AAAI 2024 · similarity 0.85)
- The Devil is in the Details: StyleFeatureEditor for Detail-Rich StyleGAN Inversion and High Quality Image Editing (CVPR 2024 · similarity 0.85)
- Preserve or Modify? Context-Aware Evaluation for Balancing Preservation and Modification in Text-Guided Image Editing (CVPR 2025 · similarity 0.85)
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.