SeamEdit: A Black-Box VLM-Agnostic Pipeline for Large-Image Semantic Editing

2026-06-11 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

SeamEdit is a novel, training-free, and model-agnostic pipeline designed for semantic region editing of large images, addressing common failure modes encountered when applying closed-source Vision-Language Models (VLMs) to tiled editing. These issues include semantic deformation, canvas-level alignment drift, and visible seam artifacts. SeamEdit treats any VLM with inpainting capability as a black-box oracle, mitigating these problems through a five-stage post-hoc process. This pipeline involves overlay-based tile decomposition, black-box VLM inpainting, geometric and color-consistency correction, seam-risk-based multi-candidate ranking, and dynamic-programming curved seam fusion. The method effectively reduces seam visibility and enables semantic modification across arbitrary tile regions, ensuring high generative quality and natural integration with surrounding content.

Key takeaway

For computer vision engineers developing large-image editing applications, SeamEdit offers a robust solution to integrate powerful black-box VLMs without incurring common tiling artifacts. You can achieve high generative quality and seamless content integration by adopting its five-stage, training-free pipeline. This approach directly mitigates semantic deformation, alignment drift, and visible seams, allowing you to use closed-source models effectively for complex semantic modifications.

Key insights

SeamEdit enables high-quality, seamless semantic editing of large images using black-box VLMs, overcoming common tiling artifacts.

Principles

• Large image editing requires generative quality and natural integration.
• Black-box VLMs can be leveraged for complex image tasks.
• Tiled editing introduces specific challenges like seams and deformation.

Method

SeamEdit employs a five-stage post-hoc pipeline: overlay-based tile decomposition, VLM inpainting, geometric/color correction, multi-candidate ranking, and dynamic-programming curved seam fusion.

In practice

• Apply SeamEdit to edit large images with closed-source VLMs.
• Use dynamic programming for curved seam fusion.
• Correct geometric and color inconsistencies post-inpainting.

Topics

• Semantic Image Editing
• Vision-Language Models
• Black-Box Models
• Image Inpainting
• Seam Fusion
• Large Image Processing

Best for: Research Scientist, AI Scientist, Computer Vision Engineer

Related on AIssential

• STiTch: Semantic Transition and Transportation in Collaboration for Training-Free Zero-Shot Composed Image Retrieval — Training-free zero-shot CIR improves by refining LLM-generated captions and using set-to-set alignment for diverse compositions.
• View-Consistent 3D Scene Editing via Dual-Path Structural Correspondense and Semantic Continuity — Achieving view-consistent 3D scene editing requires joint distribution modeling across viewpoints.
• Vision Transformer-Conditioned UNet for Domain-Adaptive Semantic Segmentation — ViTC-UNet combines frozen ViT global priors with UNet local inductive biases for efficient, high-precision biomedical semantic segmentation.
• Prompt-Guided Image Editing with Masked Logit Nudging in Visual Autoregressive Models — Masked Logit Nudging enables efficient, high-quality prompt-guided image editing in visual autoregressive models.
• VDE Bench: Evaluating The Capability of Image Editing Models to Modify Visual Documents — Image editing models require specialized benchmarks to accurately assess their performance on multilingual and densely textual visual documents.
• Boosting Zero-Shot 3D Style Transfer with 2D Pre-trained Priors — Integrating 2D pre-trained decoders into 3D style transfer pipelines effectively mitigates data scarcity and enhances stylization quality.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.