Research map / Image generation / Editing and style transfer
Prune and Repaint: Content-Aware Image Retargeting for any Ratio
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 PruneRepaint, a content-aware image retargeting method that combines saliency-guided seam carving with adaptive diffusion-based repainting of artifact-prone regions, evaluated on RetargetMe.
Problem
Image retargeting to arbitrary aspect ratios must preserve key semantics while avoiding deformation, content loss, and local artifacts such as discontinuous pixels and inconsistent regenerated content.
Approach
Combines saliency-weighted seam carving (scene- and object-level pruning via saliency and spatial centroid priors) with an adaptive repainting step that chooses inpainting or outpainting regions from pruned-pixel density and the foreground-to-target-ratio relation, using an image-guided Stable Diffusion ControlNet-inpaint with IP-Adapter.
Key results
- On RetargetMe at 16:9, SDR of 0.151 for PruneRepaint vs 0.490 for seam-carving and 0.386 for crop (lower is better).
- User study average score 2.582 at 16:9 vs 2.370 for crop and 1.720 for scale (0-3 scale, 20 volunteers).
- Ablation at 16:9: CSC alone reduces SDR from 0.490 to 0.190; adding AR reduces it to 0.151.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add content-aware seam-carving (CSC) (ablation) | Seam-carving · RetargetMe 16:9 | SDR: 0.490 -> 0.190 | Helped |
| add adaptive repainting (AR) after CSC (ablation) | CSC retargeting · RetargetMe 16:9 | SDR: 0.190 -> 0.151 | Helped |
| use adaptive repainting instead of background repainting (ablation) | CSC retargeting · RetargetMe 16:9 | SDR: 0.190 -> 0.151 | Helped |
| use PruneRepaint instead of Crop | Image retargeting · RetargetMe 16:9 | SDR: 0.386 -> 0.151 | Helped |
| use PruneRepaint instead of Crop | Image retargeting · RetargetMe 4:3 | SDR: 0.259 -> 0.074 | Helped |
| use PruneRepaint instead of Crop | Image retargeting · RetargetMe 1:1 | SDR: 0.129 -> 0.031 | Helped |
| use PruneRepaint instead of Crop | Image retargeting · RetargetMe 9:16 | SDR: 0.094 -> 0.006 | Helped |
Surprising findings
- Crop achieves the best local smoothness and deformation scores in the user study, outperforming the proposed method on those two criteria.
- Scaling scores highest on content completeness subjectively despite poor objective SDR, which the authors attribute to human interpolation ability.
Limitations
- Slow inference (about 7 s per image) due to Stable Diffusion.
- ARRD repainting regions are incomplete because it lacks global understanding, leaving misaligned structures such as streetlights.
- Evaluation limited to 80 images of RetargetMe and four aspect ratios; user study relies on 20 volunteers.
Experimental setup
Datasets: RetargetMe
Benchmarks: RetargetMe
Metrics: Saliency Discard Ratio (SDR) user study scores (content completeness, deformation, local
Baselines: Scale Crop Seam-carving InGAN Full repainting (FR) Background repainting
Builds on: seam carving Stable Diffusion ControlNet IP-Adapter InGAN
Closest papers by meaning
- VIRES: Video Instance Repainting via Sketch and Text Guided Generation (CVPR 2025 · similarity 0.86)
- Training-and-Prompt-Free General Painterly Harmonization via Zero-Shot Disentenglement on Style and Content References (AAAI 2025 · similarity 0.84)
- Infinite-Canvas: Higher-Resolution Video Outpainting with Extensive Content Generation (AAAI 2025 · similarity 0.84)
- Bricker to BRACE: A Bracket Exposure RAW Dataset and Restoration Model for Flicker-Banding (NeurIPS 2026 · similarity 0.84)
- IntrinsiX: High-Quality PBR Generation using Image Priors (NeurIPS 2025 · similarity 0.84)
- UniLumos: Fast and Unified Image and Video Relighting with Physics-Plausible Feedback (NeurIPS 2025 · similarity 0.83)
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.