Research map / Image generation / Image restoration
Reconstructing the Image Stitching Pipeline: Integrating Fusion and Rectangling into a Unified Inpainting Model
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 SRStitcher, which unifies image fusion and rectangling as a single inpainting task solved by a pre-trained Stable Diffusion inpainting model with weighted masks, requiring no training or fine-tuning.
Problem
Deep learning image stitching pipelines (registration, fusion, rectangling) are cascaded and trained separately, so registration errors propagate and degrade fusion and rectangling quality.
Approach
Reformulate fusion (seam inpainting over a coarse fusion image) and rectangling (filling missing regions) as one inpainting problem, and control per-region inpainting intensity during the diffusion reverse process with weighted masks derived from Telea coarse rectangling and distance transforms, in a single inference without prompts.
Key results
- SRStitcher reaches HIQA 47.82, CLIPIQA 33.25, and CCS 91.15% on UDIS-D test, the best among compared methods.
- On UDIS-D train, SRStitcher achieves HIQA 54.74, CLIPIQA 37.52, and CCS 93.29%, outperforming the UDIS++ + SD2 baseline (CCS 91.18%).
- A user study with 20 participants was used to compare SRStitcher against baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the baseline stitching pipeline with unified inpainting and weighted-mask guidance | Image stitching on UDIS-D test · UDIS-D test | HIQA: 46.98 -> 47.82 | Helped |
| Replace the baseline stitching pipeline with unified inpainting and weighted-mask guidance | Image stitching on UDIS-D test · UDIS-D test | CLIPIQA: 31.48 -> 33.25 | Helped |
| Replace the baseline stitching pipeline with unified inpainting and weighted-mask guidance | Image stitching on UDIS-D test · UDIS-D test | CCS (%): 89.37 -> 91.15 | Helped |
| Replace the baseline stitching pipeline with unified inpainting and weighted-mask guidance | Image stitching on UDIS-D train · UDIS-D train | HIQA: 51.49 -> 54.74 | Helped |
| Replace the baseline stitching pipeline with unified inpainting and weighted-mask guidance | Image stitching on UDIS-D train · UDIS-D train | CLIPIQA: 34.31 -> 37.52 | Helped |
| Replace the baseline stitching pipeline with unified inpainting and weighted-mask guidance | Image stitching on UDIS-D train · UDIS-D train | CCS (%): 91.18 -> 93.29 | Helped |
| Remove coarse rectangling while retaining both weighted masks (ablation) | SRStitcher · Image stitching examples | filling completeness | Hurt |
| Remove coarse rectangling and replace the weighted initial mask with the content mask (ablation) | SRStitcher · Image stitching examples | content fidelity | Hurt |
| Remove coarse rectangling and both weighted masks (ablation) | SRStitcher · Image stitching examples | abnormal-content occurrence | Hurt |
| Decrease lambda to widen the seam modification region (ablation) | SRStitcher · Stitched images | CCS | Hurt |
| Increase Telea radius R for coarse rectangling (ablation) | SRStitcher · Stitched images | weak-prior quality | Helped |
| Increase seam inpainting strength epsilon_1 from 64 to 128 (ablation) | SRStitcher · Pillar scene | shape plausibility: 64 -> 128 | Helped |
Surprising findings
- Diffusion inpainting baselines (SD1.5, SD2) produce completely different abnormal content under different seeds, whereas SRStitcher is more stable across seeds.
Limitations
- Visible seams remain when input images have significant color differences.
- Local blurring in challenging scenes due to coarse rectangling prior.
- Registration stage still relies on UDIS++ homography; not integrated.
Experimental setup
Datasets: UDIS-D
Benchmarks: UDIS-D test UDIS-D train
Metrics: HIQA CLIPIQA CCS (content consistency score)
Baselines: UDIS+DR UDISplus+DR UDIS+Lama UDISplus+Lama UDIS+SD1.5 UDISplus+SD1.5 UDIS+SD2 UDISplus+SD2
Builds on: Stable Diffusion Inpainting UDIS++ Telea inpainting Differential Diffusion
Closest papers by meaning
- Depth-Supervised Fusion Network for Seamless-Free Image Stitching (NeurIPS 2025 · similarity 0.87)
- SyFormer: Structure-Guided Synergism Transformer for Large-Portion Image Inpainting (AAAI 2024 · similarity 0.86)
- Towards Robust Image Stitching: An Adaptive Resistance Learning against Compatible Attacks (AAAI 2024 · similarity 0.86)
- TurboFill: Adapting Few-step Text-to-image Model for Fast Image Inpainting (CVPR 2025 · similarity 0.86)
- RecDiffusion: Rectangling for Image Stitching with Diffusion Models (CVPR 2024 · similarity 0.86)
- Towards Enhanced Image Inpainting: Mitigating Unwanted Object Insertion and Preserving Color Consistency (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.