Skip to content

๐ŸŽจ Image Generation

๐Ÿ’ฌ EMNLP2026 ยท 2 paper notes

๐Ÿ“Œ Same area in other venues: ๐Ÿง  NeurIPS2026 (206) ยท ๐ŸŽž๏ธ ECCV2026 (335) ยท ๐Ÿ“ท CVPR2026 (495) ยท ๐Ÿ”ฌ ICLR2026 (355) ยท ๐Ÿ’ฌ ACL2026 (5) ยท ๐Ÿงช ICML2026 (141)

Decomposing Visual Histories with Vision-Language Agents: Hierarchical Temporal Guidance for Compositional Image Generation

HTGF is a training-free framework for multi-reference compositional image generation that decomposes visual references into Subject, Structure, and Detail axes via a VLM agent, schedules their injection into SNR-aligned denoising windows derived from closed-form Tweedie sensitivity, and eliminates stage-transition artifacts using a manifold-aware corrector.

SpatialGuard: Harness-Guided Verifiable Spatial Reasoning for Text-to-Image Generation

SpatialGuard reframes complex 3D spatial text-to-image generation into an editable, verifiable layout state and introduces an agent-based Layout Harness (rule constraints, tool invocation, shared knowledge, and feedback loops) that orchestrates planning, visual realization, structured validation, and atomic repair to conquer multi-turn constraint decay, achieving a state-of-the-art 9.37 overall spatial faithfulness score.