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

Zhengyang Sun; Yu Chen; Xin Zhou; Xiaofan Li; Xiwu Chen; Dingkang Liang; Xiang Bai

cs.CVApr 9, 2026

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

Authors: Zhengyang Sun, Yu Chen, Xin Zhou, Xiaofan Li, Xiwu Chen, Dingkang Liang, Xiang Bai

Organizations: Huazhong University of Science and Technology · Afari Intelligent Drive · Zhejiang University

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.

Figures & tables

Figure 1: We present NUMINA , a training-free framework that alleviates the misalignment between precise numerals and visual instances in text-to-video diffusion models. We significantly improve counting accuracy while maintaining natural layouts and temporal coherence.

Figure 2: Visualization of the cross-attention maps corresponding to different texts in the prompt. The highlighted areas represent a stronger level of attention between the pixels and the text.

Figure 3: The pipeline of our NUMINA follows a two-phase paradigm. Given a text prompt containing numerals, we first perform the numerical misalignment identification to extract explicitly countable layouts from attention maps. Based on the layout, we further conduct a refinement and a layout-guided generation for the numerically aligned video generation.

Figure 4: The PCA visualization of self-attention maps for Wan2.1-1.3B. (a) Different attention heads naturally capture diverse spatial patterns. (b) We select the head with the highest instance separability for countable layout construction.

Models	CountAcc (%)	TC (%)	CLIP Score
Wan2.1-1.3B [ 59 ] (81 frames, 832 $\times$ 480)
Wan2.1-1.3B	42.3	81.2	33.9
+ Seed search	45.5( +3.2 )	82.3( +1.1 )	34.6( +0.7 )
+ Prompt enhancement	47.2( +4.9 )	82.1( +0.9 )	33.7( -0.2 )
+ NUMINA ( ours )	49.7 ( +7.4 )	83.4 ( +2.2 )	35.6 ( +1.7 )
Wan2.2-5B [ 59 ] (81 frames, 1280 $\times$ 704)

Table 1: Comparison of NUMINA with other practical strategies. We report Counting Accuracy (CountAcc), Temporal Consistency (TC), and CLIP Score on Wan [ 59 ] of varying scales.

Figure 5: Qualitative comparison of NUMINA with the most advanced commercial models.

Figure 6: The per-numeral accuracies for Wan2.1-1.3B.

Figure 7: Ablation on the reference timesteps

t^{\star}

for head selection.

Method	CountAcc (%)	TC (%)
Baseline	42.3	81.2
GroundingDINO	47.5( +5.2 )	82.8( +1.6 )
Attention ( ours )	49.7 ( +7.4 )	83.4 ( +2.2 )

Table 2: Ablation on the layout construction method.

$\mathcal{C}_{o}$	$\mathcal{C}_{c}$	$\mathcal{C}_{t}$	CountAcc (%)	TC (%)
Baseline			42.3	81.2
$\checkmark$			45.1( +2.8 )	82.1( +0.9 )
$\checkmark$	$\checkmark$		46.9( +4.6 )	82.3( +1.1 )
$\checkmark$		$\checkmark$	48.9( +6.6 )	83.1( +1.9 )
$\checkmark$	$\checkmark$	$\checkmark$	49.7 ( +7.4 )	83.4 ( +2.2 )

Table 3: Ablation on the components of the layout refinement cost.

Method	CountAcc (%)	TC (%)
Baseline	42.3	81.2
Random	44.1( +1.8 )	82.6( +1.4 )
All average	43.0( +0.7 )	82.4( +1.2 )
Top-3	48.2( +5.9 )	82.5( +1.3 )
Top-2	49.4( +7.1 )	83.3( +2.1 )
Top-1	49.7 ( +7.4 )	83.4 ( +2.2 )

Table 4: Ablation on the self-attention head selection strategy.

Method	wall-clock (s)	VRAM (GB)	CountAcc (%)
Wan2.1-1.3B	292	14.3	42.3
NUMINA	431	16.3	49.7
NUMINA + EasyCache [ 73 ]	355	16.3	49.4

Table 5: Additional time and VRAM cost.

Models	CountAcc (%)	TC (%)	CLIP Score
CogVideoX-5B [ 70 ] (81 frames, 1360 $\times$ 768)
CogVideoX-5B	40.2	78.1	34.8
+ Seed search	42.7( +2.5 )	78.3( +0.2 )	34.8( -0.0 )
+ Prompt enhancement	42.5( +2.3 )	79.0( +0.9 )	34.5( -0.3 )
+ NUMINA ( ours )	44.4 ( +4.2 )	80.2 ( +2.1 )	35.4 ( +0.6 )

Table 6: Evaluation results on CogVideoX [ 70 ] .

Models	CountAcc (%)	TC (%)	CLIP Score
Wan2.1-1.3B [ 59 ] (81 frames, 832 $\times$ 480)
Wan2.1-1.3B	42.3	81.2	33.9
+ Seed search	45.5( +3.2 )	82.3( +1.1 )	34.6( +0.7 )
+ Prompt enhancement	47.2( +4.9 )	82.1( +0.9 )	33.7( -0.2 )
+ NUMINA ( ours )	49.7( +7.4 )	83.4( +2.2 )	35.6 ( +1.7 )
+ Combined method ( ours )	54.2 ( +11.9 )	83.6 ( +2.4 )	35.5( +1.6 )

Table 7: Ablation on combined methods.

Models	Baseline	+ NUMINA ( ours )
Wan2.1-1.3B	83.1	83.6 ( +0.5 )
Wan2.1-14B	84.3	84.7 ( +0.4 )
Wan2.2-5B	83.4	83.5 ( +0.1 )
CogVideoX-5B	84.6	84.6 ( +0.0 )

Table 8: VBench [ 25 ] Subject-Consistency scores.

Method	CountAcc (%)	TC (%)
Baseline	42.3	81.2
No intervention	48.8( +6.5 )	83.0( +1.8 )
Rectangle	49.5( +7.2 )	83.3( +2.1 )
Circle	49.7 ( +7.4 )	83.4 ( +2.2 )

Table 9: Ablation on strategy for no-reference addition.

Figure 8: PCA visualization across timesteps and layers.

$\lambda$ / CountAcc (%)	$\tau$ / CountAcc (%)	$k$ / CountAcc (%)
4 / 49.3	0.1 / 48.4	0.5 / 48.2
8 / 49.7	0.2 / 49.7	0.8 / 49.7
16 / 49.5	0.3 / 49.2	1.0 / 49.2

Table 10: Ablation results for different hyperparameter values.

Addition	Removal	CountAcc (%)	TC (%)
Baseline		42.3	81.2
$\checkmark$		47.7( +5.4 )	83.0( +1.8 )
	$\checkmark$	43.8( +1.5 )	82.4( +1.2 )
$\checkmark$	$\checkmark$	49.7 ( +7.4 )	83.4 ( +2.2 )

Table 11: Ablation on object addition or removal.

Method	Imaging $\uparrow$	Aesthetic $\uparrow$
Wan2.1-1.3B	71.3%	61.5%
+NUMINA	70.9%	63.5%

Table 12: VBench Aesthetic & Imaging Quality scores.

Figure 9: A failure case of NUMINA. The parrots’ heads become decoupled from their bodies in layout construction.

Figure 10: More representative examples where our method faithfully generates the specified number of objects.

T2V Generation

Text-to-video generation creates moving visual scenes from natural-language prompts, with growing control over motion, identity, events, and style.

See all 111 papers

ATHENA: Adaptive Test-Time Steering for Improving Count Fidelity in Diffusion Models
Mar 20, 2026Mohammad Shahab Sepehri, Asal Mehradfar, Berk Tinaz +2T2I Generation T2I Diffusion Models
Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, USA
Compositional Video Generation via Inference-Time Guidance
May 14, 2026Ariel Shaulov, Eitan Shaar, Amit Edenzon +2Diffusion Model Guidance Video Diffusion Models
1Tel-Aviv University · 2Independent Researcher · 3Bar Ilan University +1
TPD: Temporal Prior Decoupling for Text-to-Video Diffusion Models
Jul 29, 2026Taewon Kang, Matthias ZwickerDiffusion Model Guidance Video Diffusion Models
University of Maryland at College Park, United States

Mar 20, 2026J/K move · Enter open · S save

ATHENA: Adaptive Test-Time Steering for Improving Count Fidelity in Diffusion Models

Mohammad Shahab Sepehri, Asal Mehradfar, Berk Tinaz, Salman Avestimehr+1

Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, USA

T2I Generation T2I Diffusion Models Diffusion Model Guidance Test-Time Adaptation

Open paper AskPDF arXiv

Text-to-image diffusion models achieve high visual fidelity but surprisingly exhibit systematic failures in numerical control when prompts specify explicit object counts. To address this limitation, we introduce ATHENA, a model-agnostic, test-time adaptive steering framework that improves object count fidelity without modifying model architectures or requiring retraining. ATHENA leverages intermediate representations during sampling to estimate object counts and applies count-aware noise corrections early in the denoising process, steering the generation trajectory before structural errors become difficult to revise. We present three progressively more advanced variants of ATHENA that trade additional computation for improved numerical accuracy, ranging from static prompt-based steering to dynamically adjusted count-aware control. Experiments on established benchmarks and a new visually and semantically complex dataset show that ATHENA consistently improves count fidelity, particularly at higher target counts, while maintaining favorable accuracy-runtime trade-offs across multiple diffusion backbones.

Abstract

T2V Generation

Explore similar work

ATHENA: Adaptive Test-Time Steering for Improving Count Fidelity in Diffusion Models

Compositional Video Generation via Inference-Time Guidance

TPD: Temporal Prior Decoupling for Text-to-Video Diffusion Models