Researchers have developed OcclusionFormer, a new framework designed to improve layout-grounded image generation by explicitly handling inter-object occlusion. Existing models struggle when bounding boxes overlap, leading to ambiguous or inconsistent layering. OcclusionFormer addresses this by using a novel Diffusion Transformer that models Z-order priority and employs volume rendering for compositing. The approach is supported by a new dataset, SA-Z, which includes explicit occlusion ordering and pixel-level annotations, leading to enhanced semantic consistency and accuracy in generated images. AI
影响 Improves spatial controllability in image generation models by resolving complex occlusion relationships.
排序理由 The cluster describes a new research paper introducing a novel framework and dataset for image generation.
在 Hugging Face Daily Papers 阅读 →
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →