Hallo4D: Multi-Modal Hallucination Mitigation for Consistent Spatio-Temporal Generation

Hallo4D presents a model-agnostic, training-free framework for reducing spatial and temporal hallucinations in 3D and 4D generation. Its generation-detection-correction pipeline uses large multimodal models to inspect multi-view and multi-frame renderings, summarize duplicated structures, geometric misalignment, jitter, identity flicker, and structural drift, then guide consensus-driven image-space optimization. The framework adds multi-model voting for candidate correction selection, motion-aware keyframe sampling, LMM-guided initialization, appearance alignment, exposure-aware optimization, and visibility pruning. The paper reports improvements over strong baselines across diverse 3D and 4D settings, although the supplied abstract does not provide numerical results.