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

Paper Overview

Field: Computer Vision (CV) Authors: Zhengyang Sun, Yu Chen, Xin Zhou Published: 2025-04-10 arXiv: 2504.07941

Summary

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. The authors introduce NUMINA, a training-free *identify-then-guide* framework for improved numerical alignment.

How It Works

• Identify: NUMINA detects prompt-layout inconsistencies by selecting discriminative self-attention and cross-attention heads to derive a countable latent layout.
• Guide: It then conservatively refines this layout and modulates cross-attention to guide video regeneration — all without any model training or fine-tuning.

Results

Evaluated on the introduced CountBench benchmark:

• Counting accuracy improved by up to 7.4% on the Wan2.1-1.3B model
• Improvements of 4.9% and 5.5% on 5B and 14B models, respectively
• CLIP alignment improved while temporal consistency is maintained

These results demonstrate that structural guidance complements seed search and prompt enhancement techniques, offering a practical, training-free path toward count-accurate text-to-video diffusion generation.

--- *Links: arXiv page*