The year/Independent research

Paper 2604.08546

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

Published
Apr 2026
Research lab
Independent
Citations
0
GitHub
68 stars

01 In brief

Summary

The paper introduces NUMINA, a training-free framework for improving numerical alignment in text-to-video diffusion models, which often fail to generate the correct number of objects specified in prompts.

NUMINA uses an identify-then-guide paradigm: first, it identifies prompt-layout inconsistencies by selecting discriminative self- and cross-attention heads to derive a countable latent layout; then, it refines this layout conservatively and modulates cross-attention to guide regeneration.

On the introduced CountBench benchmark (210 prompts, counts 1-8, 1-3 object categories), NUMINA improves counting accuracy by 7.4% on Wan2.1-1.3B, 4.9% on Wan2.2-5B, and 5.5% on Wan2.1-14B, while also improving CLIP alignment and maintaining temporal consistency.

The method is compatible with inference acceleration like EasyCache and generalizes to CogVideoX-5B.

Ablations show the importance of head selection, layout refinement costs, and reference timestep (t*=20).

Limitations include challenges with dense instances and over-segmentation of salient object parts.

02 From the paper

Abstract

Text-to-video diffusion models have enabled open-ended video synthesis, but often struggle with generating the correct number of objects specified in a prompt. We introduce NUMINA , a training-free identify-then-guide framework for improved numerical alignment. NUMINA identifies prompt-layout inconsistencies by selecting discriminative self- and cross-attention heads to derive a countable latent layout. It then refines this layout conservatively and modulates cross-attention to guide regeneration. On the introduced CountBench, 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. Furthermore, CLIP alignment is improved while maintaining temporal consistency. These results demonstrate that structural guidance complements seed search and prompt enhancement, offering a practical path toward count-accurate text-to-video diffusion. The code is available at https://github.com/H-EmbodVis/NUMINA.