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

SeamEdit is a post-hoc pipeline for semantic editing of large images that does not require access to model internals. Instead of relying on white-box methods, it treats any vision-language model with inpainting support as a black-box oracle, which is useful when the strongest generation models are closed-source.

The paper focuses on a practical pain point: tiled editing of large images often breaks down at tile boundaries. The authors call out three common failures—semantic deformation, canvas-level alignment drift, and visible seam artifacts—and propose a five-stage fix: 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.

For game developers, this is most relevant to tools, tech art, and content pipelines that need high-quality large-image edits such as concept art cleanup, marketing key art, texture/paintover workflows, or UI image manipulation. The main appeal is that it is training-free and model-agnostic, so it can potentially slot around existing proprietary models instead of forcing a custom training stack.

This is still an arXiv paper, submitted on 11 Jun 2026, so the real question is whether the seam-fusion and correction stages hold up across messy production assets. If it does, it points toward a useful pattern: keep the generative model as a black box, then do the hard production work in deterministic post-processing.