When Numbers Speak: Aligning Textual Numerals and Visual Instances in Text-to-Video Diffusion Models

NUMINA is a training-free framework that tackles numerical misalignment in text-to-video diffusion models — the persistent failure of T2V models to generate the correct count of objects specified in prompts (e.g., producing 2 or 4 cats when "three cats" is requested). Unlike seed search or prompt enhancement approaches that treat the generation pipeline as a black box and rely on brute-force resampling or LLM-based prompt rewriting, NUMINA directly identifies where and why counting errors occur inside the model by analyzing cross-attention and self-attention maps at selected DiT layers. It constructs a countable spatial layout via a two-stage clustering pipeline, then performs layout-guided attention modulation during regeneration to enforce the correct object count — all without retraining or fine-tuning. On our introduced CountBench, this attention-level intervention provides principled, interpretable control over numerical semantics that seed search and prompt enhancement fundamentally cannot achieve, improves counting accuracy by up to 7.4% on Wan2.1-1.3B. Furthermore, because NUMINA operates partly orthogonally to inference acceleration techniques, it is compatible with training-free caching methods such as EasyCache, which accelerates diffusion inference via runtime-adaptive transformer output reuse.