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.
Figures & tables
Figure 1: Count fidelity failures in T2I generative models. Despite explicit numerical specifications in the prompts, generated images frequently contain incorrect object counts across multiple generative backbones, even for simple objects and scenes, highlighting a persistent limitation in numerical control.
Figure 2: Count estimation across diffusion steps. At early sampling steps, object structure is weakly formed and decoded images are too noisy for reliable count estimation. At later steps, object instances become well defined, and counting is reliable, but requires additional diffusion steps and a higher inference cost. This trade-off motivates estimating object count at an intermediate diffusion step.
Figure 3: ATHENA steering block. The denoiser is evaluated with the original and control prompts to obtain ϵt and ϵ^t , which are combined to obtain the latent state zt−1 for test-time count control.
Figure 4: ATHENA-Adaptive pipeline. The method estimates object count at an intermediate diffusion step, applies early-stage prompt-based steering, adaptively adjusts the steering strength once based on the observed error direction, and completes generation using standard diffusion.
Level
Description
Example Prompt
L1
Hard object categories
“nine microphones”
L2
+ Scene context
“eight sneakers at the edge of a pond”
L3
+ Distractor objects
“five jars beside a river with a person”
L4
+ Relational constraints
“seven dumbbells lined up along a sidewalk”
Table 1: Structure of the ATHENA dataset with four levels of increasing prompt complexity.
Model
Method
CoCoCount
CoCoCount-E
ATHENA Dataset
Acc ( ↑ )
MAE ( ↓ )
RMSE ( ↓ )
Time ( ↓ )
Acc ( ↑ )
MAE ( ↓ )
RMSE ( ↓ )
Time ( ↓ )
Acc ( ↑ )
MAE ( ↓ )
RMSE ( ↓ )
Time ( ↓ )
FLUX.1-dev
Unsteered
58.4
0.98
1.96
45.8
46.5
1.12
1.93
22.6
39.4
1.78
2.96
45.5
ATHENA-Static
73.3
0.56
1.35
54.8
58.5
0.84
1.73
27.4
48.9
1.31
2.44
54.7
ATHENA-Feedback
71.4
0.67
1.72
56.0
58.3
0.98
2.00
29.8
52.2
1.38
2.60
60.8
ATHENA-Adaptive
70.2
0.65
1.63
64.8
62.2
0.85
1.90
35.1
53.6
1.40
2.72
72.5
SD 3.5 Large
Unsteered
58.4
0.78
1.50
43.9
44.8
0.98
1.56
24.1
38.9
1.55
2.54
43.8
Table 2: Quantitative counting performance across diffusion backbones and datasets. Accuracy (%) is reported as exact-match count accuracy, with the best result for each model–dataset pair shown in bold . MAE and RMSE measure counting error, and Time denotes mean generation time per sample (seconds).
Figure 7
Figure 7: Qualitative results on complex, distractor-rich prompts. Guidance denotes Counting Guidance. Green borders indicate correct counts; red borders indicate incorrect counts. All ATHENA variants preserve scene structure and color consistency while handling relational and multi-object instructions. Compared to prior baselines, ATHENA avoids visual artifacts and background distortion, with adaptive steering yielding the most accurate counts.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: ATHENA-Static pipeline. Starting from the same initial noise, prompt-based steering is applied using a fixed, count-agnostic control prompt from diffusion step T down to a cutoff tsteer . Standard diffusion then proceeds unmodified to completion, yielding improved count fidelity without intermediate feedback or adaptive adjustment.
Figure 9: ATHENA-Feedback pipeline. An intermediate count is estimated at step test without steering. If the estimate differs from the target, generation is restarted from the same initial noise using a feedback control prompt, and a single early-stage steering phase is applied before completing diffusion.
Figure 10: ATHENA-Adaptive pipeline. The method estimates object count at an intermediate diffusion step, applies early-stage prompt-based steering, adaptively adjusts the steering strength once based on the observed error direction, and completes generation using standard diffusion.
Dataset
Unique Prompts
Count Range
Total Prompts
CoCoCount (deduplicated)
161
2–10
161
CoCoCount-E
72
2–10
648
Appendix
Table 3: Dataset statistics for CoCoCount and CoCoCount-E.
Diffusion Backbone
Parameter
FLUX.1-dev
SD 3.5 Large
SDXL
Static
Feedback
Adaptive
Static
Feedback
Adaptive
Static
Feedback
Adaptive
test
–
20
20
–
20
20
–
30
30
tsteer
10
5
5
10
5
5
10
10
10
γ
4
4
4
4
4
4
5
5
5
Appendix
Table 4: Hyperparameter settings used for ATHENA across diffusion backbones and steering variants.
Figure 11: Accuracy–runtime trade-offs on CoCoCount and CoCoCount-E. Colors denote the diffusion backbone, while marker shapes indicate the method. ATHENA-Adaptive consistently achieves higher accuracy at comparable or lower runtime than prior baselines.
Figure 12: Counting accuracy versus target object count across datasets and diffusion backbones. ATHENA-Adaptive consistently maintains higher accuracy as the target count increases, demonstrating improved robustness to increasing counting difficulty.
Figure 13: Additional qualitative results on relational and multi-object prompts. Results are shown for unsteered generation (FLUX) and the three ATHENA variants. Green borders indicate correct object counts, while red borders denote counting errors. ATHENA improves count fidelity across diverse settings (e.g., grouping, circular layouts, and natural scenes) while preserving scene structure and visual coherence, with the adaptive variant yielding the most consistent corrections.
Figure 14: Additional qualitative results on relational and multi-object prompts. Results are shown for unsteered generation (SD 3.5) and the three ATHENA variants. Green borders indicate correct object counts, while red borders denote counting errors. ATHENA improves count fidelity across diverse settings (e.g., grouping, circular layouts, and natural scenes) while preserving scene structure and visual coherence, with the adaptive variant yielding the most consistent corrections.
Figure 15: Additional qualitative results on relational and multi-object prompts with the SDXL backbone. Results include unsteered generation, CountGen, Counting Guidance ( Guidance ), and the three ATHENA variants. Green borders indicate correct object counts, while red borders denote counting errors. Prior baselines frequently fail to enforce correct counts or introduce visual artifacts, whereas ATHENA improves count fidelity while preserving scene structure and visual coherence, with the adaptive variant yielding the most consistent corrections.
Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.
Diffusion models excel at photorealistic synthesis but struggle with object count fidelity, especially in high-density settings. We introduce COUNTLOOP, a training-free framework that achieves structured instance and count control through iterative, structured feedback. Our method alternates between synthesis and evaluation: a VLM-based planner generates structured scene layouts, while a VLM-based critic provides explicit feedback on object counts, spatial arrangements, and visual quality to refine the layout iteratively. Instance-driven attention masking and cumulative attention composition further prevent semantic leakage, ensuring clear object separation even in densely occluded scenes. Evaluations on COCO-Count, T2I-CompBench, and two newly introduced high instance benchmarks show that COUNTLOOP reduces counting error by up to 57% and achieves the highest or comparable spatial quality scores across all benchmarks, while maintaining photorealism. Project page is at https://mondalanindya.github.io/CountLoop/.
Anindya Mondal, Sauradip Nag, Ayan Banerjee +3
University of Surrey · Simon Fraser University · Universitat Autònoma de Barcelona
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.
Zhengyang Sun, Yu Chen, Xin Zhou +4
Huazhong University of Science and Technology · Afari Intelligent Drive · Zhejiang University
Text-to-Image (T2I) diffusion/flow models have recently achieved remarkable progress in visual fidelity and text alignment. However, they remain limited when users need to precisely control image layouts, something that natural language alone cannot reliably express. Controllable generation methods augment the initial T2I model with additional conditions that more easily describe the scene. Prior works straightforwardly train the augmented network with the same loss as the initial network. Although natural at first glance, this can lead to very long training times in some cases before convergence. In this work, we revisit the training objective of controllable diffusion models through a detailed analysis of their denoising dynamics. We show that direct supervision on the clean target image, dubbed x0-supervision, or an equivalent re-weighting of the diffusion loss, yields faster convergence. Experiments on multiple control settings demonstrate that our formulation accelerates convergence by up to 2× according to our novel metric (mean Area Under the Convergence Curve - mAUCC), while also improving both visual quality and conditioning accuracy. Our code is available at https://github.com/CEA-LIST/x0-supervision
Amadou S. Sangare, Adrien Maglo, Mohamed Chaouch +1
Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France