NUMINA: Training-Free Numerical Alignment for Text-to-Video Diffusion Models

Summary

NUMINA is a training-free identify-then-guide framework that improves numerical alignment in text-to-video diffusion models, addressing their common failure to generate the correct number of objects specified in prompts. The method identifies prompt-layout inconsistencies by selecting discriminative self- and cross-attention heads to derive a countable latent layout, then conservatively refines this layout and modulates cross-attention to guide regeneration. Evaluated on the newly introduced CountBench benchmark, NUMINA improves counting accuracy by up to 7.4% on Wan2.1-1.3B, and by 4.9% and 5.5% on 5B and 14B models, respectively. CLIP alignment also improves while temporal consistency is maintained. The results, from arXiv paper 2504.07083 by Zhengyang Sun, Yu Chen, and Xin Zhou, demonstrate that structural guidance complements seed search and prompt enhancement, offering a practical path toward count-accurate text-to-video generation.