acceptodds
Withdrawn from ICLR 2027

ATHENA: Test-Time Steering for Count Fidelity in Text-to-Image Diffusion Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Text-to-image diffusion models achieve high visual fidelity yet exhibit systematic failures when prompts specify explicit object counts. We introduce ATHENA, a model-agnostic, training-free framework for improving count fidelity through test-time steering. ATHENA constructs a cardinality-specific contrastive steering signal by varying the numerical condition in the input prompt, enabling targeted count intervention during denoising while preserving the underlying semantic content. However, different generation trajectories develop different count errors during denoising, making a fixed steering signal insufficient. To address this, we further incorporate intermediate count feedback to identify emerging cardinality errors, construct trajectory-specific corrections, and adjust the steering strength based on the observed response. Across multiple text-to-image diffusion backbones and counting benchmarks, we show that ATHENA consistently improves count fidelity over unsteered generation and existing counting methods, with particularly strong gains in challenging counting settings while maintaining favorable accuracy-runtime trade-offs.

Primary area:
generative models
Keywords:
Text-to-Image Generation, Diffusion Models, Flow Matching, Test-Time Steering, Object Counting, Count Fidelity, Numerical Control

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HJs4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ATHENA, a training-free framework for improving object-count fidelity in text-to-image generation. It constructs cardinality-specific steering signals by contrasting model predictions under the original prompt and a count-modified control prompt. Three strategies are introduced: Static, Feedback, and Adaptive, which progressively incorporate intermediate count estimation and steering-strength adjustment.

The authors provide a geometric interpretation and introduce ATHENA-COUNT, a benchmark with 360 prompts of varying complexity. Experiments across several generative backbones demonstrate substantial improvements in count accuracy, supported by additional evaluations using Qwen3-VL, GenEval, and multiple random seeds.

Critical strengths

S1. Substantial empirical improvements. ATHENA improves count fidelity across multiple generative backbones and counting benchmarks. The independent Qwen3-VL evaluation, GenEval experiments, and multi-seed analysis provide additional evidence of practical effectiveness.

S2. Practical training-free framework. The method requires no retraining or architectural modifications and provides different accuracy–runtime trade-offs. The introduced counting benchmark and multi-object experiments further broaden the evaluation.

Critical weaknesses

W1. Unresolved inconsistencies in experimental reporting. Table 2 and Appendix Table 10 contain several incompatible results under the stated single-seed and three-seed protocols. For example, SDXL unsteered accuracy on ATHENA-COUNT is 19.7% in Table 2 but 24.9 ± 1.8% in Table 10, while CountGen runtime on CoCoCount-E is 29.2 s versus 45.8 ± 0.7 s. These discrepancies require reconciliation before the reported performance and efficiency conclusions can be assessed reliably.

W2. Limited methodological isolation and theory–implementation mismatch. The steering formulation resembles existing contrastive guidance, while comparisons against computation-matched alternatives are missing. Moreover, Appendix B analyzes an unnormalized noise-prediction update, whereas Algorithm 2 applies norm matching and the evaluated backbones include flow-matching models. The current derivation therefore does not directly establish the claimed geometric behavior of the implemented method.

Questions for authors

Q1. Experimental consistency. What explains the discrepancies between Table 2 and Table 10? Please provide per-seed results, runtime measurements, and any differences in sampler settings, hardware, or evaluation protocols.

Q2. Theoretical justification. Does the geometric interpretation remain valid after the norm matching in Algorithm 2? How is the derivation adapted to the velocity predictions of flow-matching models?

Q3. Contribution attribution. How does ATHENA compare with standard contrastive guidance and detector-based best-of-N selection under matched denoiser evaluations or wall-clock budgets? What is the additional benefit of feedback-based adaptation beyond increased inference computation?

Q4. Adaptive behavior. Why does Static outperform Adaptive in some settings, particularly FLUX.1-dev on CoCoCount and FLUX.1-dev / SD 3.5 Large on GenEval? Can the authors analyze correction success rates and cases where adaptive steering degrades the result?

Q5. Evaluation robustness. Although Qwen3-VL provides independent evaluation, ATHENA-COUNT explicitly prioritizes detector-compatible categories. Can the authors provide a manually verified subset or additional analysis of detector disagreement? Also, how robust are the claims of semantic preservation given the declines in some perceptual metrics reported in Table 14?

4 comments
AuthorsOpenReview ↗

Addressing reviewer HJs4 concerns (part 1)

We thank the reviewer for their careful evaluation of our work and for recognizing the empirical improvements, practical applicability, and extensive evaluation of ATHENA. We address each concern in detail below.

W1 and Q1

We thank the reviewer for identifying these discrepancies and apologize for the inconsistencies in the reported results. We would first like to clarify that all results in Table 2 and the associated figures are accurate and remain unchanged. Table 2 reports results using seed 23, while Table 10 reports the mean and standard deviation across three seeds {7, 23, 42}. The discrepancies arose from a small issue in our merger code affecting some of the entries and a few transcription errors in Table 10. We have carefully rechecked the results and updated Table 10 in the revised manuscript. For full transparency, we provide the complete per-seed results in Section G.1.1 of the Appendix, highlighted in blue, which can be used to independently verify all entries in the corrected Table 10.

Regarding runtime measurements, as specified in Section 4.1 (Computational Setup), we use NVIDIA RTX A6000 GPUs (48 GB) for CoCoCount and ATHENA-COUNT, and NVIDIA A100 SXM4 GPUs (40 GB) for CoCoCount-E. Within each dataset, all models and methods are evaluated using the same GPU type, ensuring consistent runtime comparisons. We hope these clarifications address the reviewer's concerns.

W2 and Q2

Regarding norm matching, the geometric interpretation and norm matching serve complementary purposes. Section 3.2.2 motivates the direction of the contrastive correction under the stated assumptions. Norm matching allows us to apply this correction while preserving the original prediction scale, reducing the risk of subsequent latents becoming out of distribution for the denoiser. Without it, our preliminary experiments showed severe degradation or invalid images. Related guidance-rescaling strategies have also been used to mitigate image-quality degradation in (Lin et al., 2024). Although normalization changes the exact correction relative to the unsteered prediction, it preserves the direction of the combined steered prediction. Hence, the geometric interpretation motivates the steering signal in ATHENA, rather than providing an exact characterization of the normalized update.

Regarding flow-matching models, the geometric interpretation also applies to linear flow matching. Consider zt=(1−t)z0+tξz_t=(1-t)z_0+t\xi, with $v_{t,q}$ approximating E[ξ−z0∣zt,q]\mathbb E[\xi-z_0\mid z_t,q]. Then the corresponding noise prediction and score satisfy

\qquad s_{t,q}(z_t)\approx-\frac{\epsilon_{t,q}^{\mathrm{noise}}}{t},$$ For the unnormalized steering update $\tilde v_t=v_{t,p}+\gamma(v_{t,p}-v_{t,\hat p})$, with steering strength $\gamma>0$, define the original and steered clean estimates as $\hat z_{0,p}=z_t-tv_{t,p}$ and $\tilde z_0=z_t-t\tilde v_t$. Their difference is $$\tilde z_0-\hat z_{0,p} =-t\gamma(v_{t,p}-v_{t,\hat p}) \approx\frac{\gamma t^2}{1-t} \bigl(s_{t,p}(z_t)-s_{t,\hat p}(z_t)\bigr).$$ Thus, the clean-estimate correction follows the same score contrast as in the diffusion formulation. Under the local manifold and control-condition assumptions in Appendix B, this motivates a correction toward the target-conditioned manifold. We have added the full derivation and its assumptions to **Appendix B (Extension to Linear Flow-Matching Models)**, highlighted in blue. **Reference** S. Lin, B. Liu, J. Li, and X. Yang. “Common Diffusion Noise Schedules and Sample Steps Are Flawed.” 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 5392–5399, 2024.
AuthorsOpenReview ↗

Addressing reviewer HJs4 concerns (part 2)

W2 and Q3

Regarding the Best-of-N comparison, following the reviewer’s suggestion, we evaluated detector-based Best-of-2 across all three backbones and datasets. This baseline generates two independent unsteered images and selects the one whose detected count is closest to the target. Complete results are provided below and in Appendix H.2 (Comparison with Best-of-2 Sampling), highlighted in blue. ATHENA-Adaptive achieves higher or comparable accuracy across all nine settings and lower wall-clock generation time in six. For example, on SD 3.5 Large and CoCoCount, ATHENA achieves 78.3% accuracy in 62.0 seconds, compared with 72.7% in 88.5 seconds for Best-of-2.

ModelMethodCoCoCount Acc. (%) ↑Time (s) ↓CoCoCount-E Acc. (%) ↑Time (s) ↓ATHENA-COUNT Acc. (%) ↑Time (s) ↓
FLUX.1-devBest-of-272.791.061.145.653.391.0
FLUX.1-devATHENA-Adaptive70.264.862.235.153.672.5
SD 3.5 LargeBest-of-272.788.562.748.353.187.9
SD 3.5 LargeATHENA-Adaptive78.362.065.638.156.170.4
SDXLBest-of-246.618.739.410.531.718.7
SDXLATHENA-Adaptive54.019.445.511.634.721.2

Beyond these results, ATHENA, like correction-based baselines, targets errors in the same seeded generation, aiming to preserve its scene structure and semantic content rather than replace it with an independently generated image. Moreover, additional feedback-guided correction attempts could further refine the steering strength γ\gamma. Resampling and steering are also complementary. Best-of-N can select among independently generated ATHENA candidates, combining ATHENA’s higher per-sample success rate with the benefits of candidate selection.

Regarding the benefit of feedback, we thank the reviewer for raising this point. Feedback allows ATHENA to construct a control prompt that better targets the actual counting discrepancy, guiding the correction toward the desired count and avoiding unnecessary intervention when the estimated count is already correct. Adaptive further adjusts the correction strength based on the observed response. This targeted approach achieves higher exact-count accuracy on the more challenging CoCoCount-E and ATHENA-COUNT datasets across all three backbones (Table 2). For example, on ATHENA-COUNT, Feedback achieves 52.2% and 51.1% accuracy on FLUX.1-dev and SD 3.5 Large, respectively, compared with 48.9% and 46.7% for Static. We would also be happy to evaluate any additional specific contrastive-guidance baseline the reviewer suggests.

Q4

We thank the reviewer for this question. We would like to clarify that ATHENA-Adaptive is particularly more effective on harder datasets; for example, on ATHENA-COUNT with FLUX.1-dev, it achieves 53.6% accuracy versus 48.9% for Static. On the other hand, Static can be more effective on simpler settings with simpler prompts and smaller target counts, as commonly encountered in CoCoCount and GenEval. In these cases, additional feedback-based adjustments may offer limited benefit or introduce errors due to imperfect intermediate count estimates. For example, Figure 9(a) shows Static outperforming Adaptive at a target count of four on CoCoCount-E with FLUX.1-dev.

Moreover, following the reviewer’s suggestion, we compared unsteered and ATHENA-Adaptive outputs on ATHENA-COUNT using FLUX.1-dev. 16.67% of all samples were incorrect without steering but correct with Adaptive, while 2.50% showed the reverse, yielding an overall accuracy gain from 39.44% to 53.61%. Appendix J.2 illustrates successful and unsuccessful corrections and their effects on scene structure.

Q5

Regarding detector compatibility and human evaluation, we additionally conducted a human evaluation of SD 3.5 Large generations on ATHENA-COUNT based on the reviewer's suggestion. Human-evaluated exact-count accuracy improves from 47.50% for Unsteered to 61.67% for ATHENA-Adaptive, consistent with the improvements reported by G-DINO (38.90% to 56.10%) and Qwen3-VL (44.44% to 60.83%, Table 3). The human evaluation is closer to Qwen3-VL than G-DINO for both methods, suggesting that G-DINO may underestimate counting accuracy on ATHENA-COUNT. Overall, this result supports that ATHENA’s gains are not specific to detector-based evaluation.

Regarding image quality, the drops in PickScore and HPSv2 are small (Appendix J.1, Table 14), suggesting that any degradation in perceived image quality is unlikely to be noticeable to human observers. The qualitative examples in Appendix J.2 also support this, showing count corrections that preserve scene structure and visual quality.

We hope the new experiments and discussion resolve the reviewer’s concerns and clarify ATHENA’s contribution and practical benefits.

Thank you for the further clarifications. I accept the explanations regarding baseline availability, backbone differences, and the range of counts. The revised experimental results and supplementary derivations have largely addressed my concerns about the paper’s reliability.

My primary remaining reservation concerns the novelty of the method and the attribution of performance gains. While the additional "Best-of-2" experiment is valuable, it does not yet sufficiently demonstrate that ATHENA offers a distinct advantage over simple guidance based on numerical constraints.

Overall, the rebuttal has significantly strengthened the credibility of the paper; however, I will maintain my current score for the time being.

We thank the reviewer for their response and are pleased that the additional experiments and clarifications have largely addressed the reliability concerns.

We would appreciate further guidance on the remaining concern. Our evaluation includes Counting Guidance, which uses gradients to enforce numerical constraints. In same-backbone comparisons, ATHENA achieves higher accuracy with substantially lower computational overhead. To ensure we address the reviewer’s intended comparison, could they clarify which method or setup they have in mind by “simple guidance based on numerical constraints”? We would be happy to conduct additional comparisons to help address this concern.

Reviewer_WVH5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper tackles the problem of improving object count accuracy in test-to-image diffusion models. The authors propose a training-free method named ATHENA, which construct a count-specific steering direction by comparing the model’s predictions under the original prompt and a modified prompt that changes only the number. They also introduce a now benchmark. Experiments on several diffusion models show the effectiveness of the proposed method.

Critical strengths
  1. The proposed method is simple and intuitive, requires no additional training or architectural modifications and can be applied to different diffusion and flow-matching models.

  2. The experiments cover multiple generative backbones and counting benchmarks, with additional analyses of generalization, robustness, computational efficiency, and image quality.

Critical weaknesses
  1. The novelty of the proposed method may be limited. The proposed steering formulation is similar to classifier-free and training-free guidance in diffusion model, which has been extensively studied in previous works. Thee main difference is the construction of count-specific prompt, but this does not sufficiently justify technical contribution. Meanwhile, although the experimental results are promising, the paper does not sufficiently compare the proposed method with existing guidance techniques or evaluate simpler alternatives under comparable inference budgets. This makes it difficult to assess whether the improvements stem from a genuinely new steering mechanism or from adapting established guidance strategies to object counting.

  2. The experimental comparison may not be complete. While the paper discusses some recent related works, including CountCluster, CountDiffusion, and CountLoop, it does not include them in the experimental comparisons. Also, some comparisons involve different generative backbones, such as comparing ATHENA on SDXL with YOLO-Count on SDXL-Turbo. Although the paper provides same-backbone comparisons with CountGen and Counting Guidance, the missing methods and backbone differences make it difficult to tell how much ATHENA actually outperforms existing approaches.

Questions for authors
  1. The geometric interpretation in Section 3.2 and Appendix B appears to analyze the unnormalized steering update, while the actual implementation normalizes the steered denoiser prediction. Could the authors clarify whether this normalization affects the theoretical justification and provide an ablation without normalization?

  2. The paper only evaluates the method on object counts ranging from 2 to 10. Higher counts, where existing methods struggle most, are not evaluated.

  3. The theory relies on strong assumptions. For example, it assumes the current image already matches the control prompt. These assumptions are not checked empirically. Also, the equations and derivations are for standard diffusion models, but FLUX and SD 3.5 are flow-matching models.

2 comments
AuthorsOpenReview ↗

Addressing reviewer WVH5 concerns (part 1)

We thank the reviewer for their thoughtful evaluation and for recognizing the simplicity, training-free nature, and broad applicability of ATHENA, as well as the extensive experimental evaluation. We address the concerns and questions below.

W1

Regarding methodological novelty, we understand the reviewer’s concern about ATHENA’s CFG-style prediction contrast, but this equation is not our main contribution. Drawing on prior findings about T2I generation, we designed intermediate count estimation, feedback-informed control prompts, and selective early-stage steering as a unified framework that balances efficiency, effectiveness, and portability. Our evaluations show that this lightweight design outperforms more complex, model-specific approaches such as CountGen and computationally demanding methods such as Counting Guidance. Our methodological contribution is not simply constructing a count-specific prompt, but combining error detection, selective intervention, and adaptive correction into a unified framework.

Regarding simpler alternatives and computational comparisons, following the reviewer’s suggestion, we evaluated detector-based Best-of-2 across all three backbones and datasets. This baseline generates two independent unsteered images and selects the one whose detected count is closest to the target. Complete results are provided below and in Appendix H.2 (Comparison with Best-of-2 Sampling), highlighted in blue. ATHENA-Adaptive achieves higher or comparable accuracy across all nine settings and lower wall-clock generation time in six. For example, on SD 3.5 Large and CoCoCount, ATHENA achieves 78.3% accuracy in 62.0 seconds, compared with 72.7% in 88.5 seconds for Best-of-2.

ModelMethodCoCoCount Acc. (%) ↑Time (s) ↓CoCoCount-E Acc. (%) ↑Time (s) ↓ATHENA-COUNT Acc. (%) ↑Time (s) ↓
FLUX.1-devBest-of-272.791.061.145.653.391.0
FLUX.1-devATHENA-Adaptive70.264.862.235.153.672.5
SD 3.5 LargeBest-of-272.788.562.748.353.187.9
SD 3.5 LargeATHENA-Adaptive78.362.065.638.156.170.4
SDXLBest-of-246.618.739.410.531.718.7
SDXLATHENA-Adaptive54.019.445.511.634.721.2

Beyond these results, ATHENA, like correction-based baselines, targets errors in the same seeded generation, aiming to preserve its scene structure and semantic content rather than replace it with an independently generated image. Moreover, additional feedback-guided correction attempts could further refine the steering strength γ\gamma. Resampling and steering are also complementary. Best-of-N can select among independently generated ATHENA candidates, combining ATHENA’s higher per-sample success rate with the benefits of candidate selection.

W2

Regarding the suggested baselines, complete implementations of CountDiffusion, CountCluster, and CountLoop were unavailable at submission and rebuttal time. CountDiffusion has no official public code, CountCluster's repository provides only a README, and CountLoop's repository states that its release remains under preparation, with key modules remaining placeholders as stated. Our attempts to reproduce some of these methods yielded low accuracy, but without reference implementations, we could not determine whether this reflected their performance or implementation discrepancies. We therefore excluded these results to avoid unfair comparisons.

Regarding backbone selection and comparison fairness, we agree that same-backbone comparisons are important. However, most existing baselines are not model-agnostic, a limitation ATHENA aims to address. We adapted Counting Guidance to FLUX.1-dev and SD 3.5 Large and evaluated ATHENA on SDXL to enable direct comparisons with Counting Guidance and CountGen, respectively. Moreover, YOLO-Count is included as a representative of gradient-based counting control in a single-step generation setting. In this method, iterative embedding optimization makes multi-step generation computationally impractical, so we retain its published SDXL-Turbo configuration, which has a runtime comparable to ATHENA on SDXL.

AuthorsOpenReview ↗

Addressing reviewer WVH5 concerns (part 2)

Q1

Regarding norm matching, the geometric interpretation and norm matching serve complementary purposes. Section 3.2.2 motivates the direction of the contrastive correction under the stated assumptions. Norm matching allows us to apply this correction while preserving the original prediction scale, reducing the risk of subsequent latents becoming out of distribution for the denoiser. Without it, our preliminary experiments showed severe degradation or invalid images. Related guidance-rescaling strategies have also been used to mitigate image-quality degradation in (Lin et al., 2024). Although normalization changes the exact correction relative to the unsteered prediction, it preserves the direction of the combined steered prediction. Hence, the geometric interpretation motivates the steering signal in ATHENA, rather than providing an exact characterization of the normalized update.

Reference

S. Lin, B. Liu, J. Li, and X. Yang. “Common Diffusion Noise Schedules and Sample Steps Are Flawed.” 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 5392–5399, 2024.

Q2

Regarding the evaluation range, prior work, including Make It Count (Binyamin et al., 2025), focuses on target counts from 2 to 10, as counting accuracy decreases substantially with increasing cardinality. Guo et al. (2025) further show that for target counts of 11–15, 12 of 15 evaluated models, including proprietary systems, achieve exact-count accuracy below 10%, with some as low as 2%. Following these findings from prior work, we focus our evaluation on the range of 2–10 objects, which remains challenging for existing text-to-image generators. Although ATHENA imposes no restriction on the target count, we agree that its effectiveness, like that of steering-based baselines, is likely to be limited at higher counts. We will explicitly acknowledge this limitation in the discussion.

References

[1] L. Binyamin, Y. Tewel, H. Segev, E. Hirsch, R. Rassin, and G. Chechik. “Make It Count: Text-to-Image Generation with an Accurate Number of Objects.” 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 13242–13251, 2025.

[2] X. Guo, J. Huo, Y. Liang, Z. Shi, Z. Song, J. Zhang, and Z. Zhuang. “Text-to-Image Diffusion Models Cannot Count, and Prompt Refinement Cannot Help.” arXiv preprint arXiv:2503.06884, 2025.

Q3

Regarding the theoretical assumptions, the geometric interpretation in Section 3.2.2 and Appendix B is intended to provide a theoretical motivation for the proposed steering mechanism, rather than a formal proof of its effectiveness. In particular, the assumption that the current latent state is locally aligned with the control-conditioned manifold is an approximation that helps explain why the contrast between model predictions can provide a meaningful count-specific steering direction. We do not claim that this assumption holds exactly for every generation trajectory. ATHENA's consistent improvements across different diffusion backbones and counting benchmarks (Table 2) provide empirical support for this geometric intuition, although these results do not directly validate the underlying assumptions. Rather, they demonstrate the practical effectiveness of the proposed steering mechanism even when the assumptions may not hold exactly.

Regarding flow-matching models, the geometric interpretation also applies to linear flow matching. Consider zt=(1−t)z0+tξz_t=(1-t)z_0+t\xi, with $v_{t,q}$ approximating E[ξ−z0∣zt,q]\mathbb E[\xi-z_0\mid z_t,q]. Then the corresponding noise prediction and score satisfy

\qquad s_{t,q}(z_t)\approx-\frac{\epsilon_{t,q}^{\mathrm{noise}}}{t},$$ For the unnormalized steering update $\tilde v_t=v_{t,p}+\gamma(v_{t,p}-v_{t,\hat p})$, with steering strength $\gamma>0$, define the original and steered clean estimates as $\hat z_{0,p}=z_t-tv_{t,p}$ and $\tilde z_0=z_t-t\tilde v_t$. Their difference is $$\tilde z_0-\hat z_{0,p} =-t\gamma(v_{t,p}-v_{t,\hat p}) \approx\frac{\gamma t^2}{1-t} \bigl(s_{t,p}(z_t)-s_{t,\hat p}(z_t)\bigr).$$ Thus, the clean-estimate correction follows the same score contrast as in the diffusion formulation. Under the local manifold and control-condition assumptions in Appendix B, this motivates a correction toward the target-conditioned manifold. We have added the full derivation and its assumptions to **Appendix B (Extension to Linear Flow-Matching Models)**, highlighted in blue. We hope these additional results and explanations address the points raised and provide a clearer picture of ATHENA’s novelty and usefulness.
Reviewer_ciiEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces ATHENA, a training-free framework for improving object-count accuracy in text-to-image diffusion and flow-matching models through test-time steering. It constructs a contrastive signal from prompts that preserve the scene description while changing the numerical specification. Its feedback-based variants use intermediate count estimates to select the control prompt and optionally adjust the steering strength. The paper also introduces ATHENA-COUNT, a benchmark incorporating challenging object categories, scene contexts, distractors, and spatial relations. Experiments across several backbones report exact-count accuracy improvements of up to approximately 22 percentage points over unsteered generation. On SDXL, ATHENA also offers favorable runtime and memory usage relative to CountGen. The reviewer suggests a clear rejection considering the limited scope and incremental novelty.

Critical strengths
  • Simple and practical implementation.

ATHENA operates through forward denoiser evaluations without backpropagation or attention manipulation. Its application across UNet and transformer-based generators demonstrates practical portability, while the reported results show favorable accuracy–runtime trade-offs against the evaluated baselines.

  • A valuable benchmark focused on counting.

ATHENA-COUNT complements simpler counting benchmarks by introducing structured difficulty levels involving challenging objects, scene context, distractors, and spatial relations. This provides a useful setting for assessing counting performance under more complex prompts.

Critical weaknesses
  • Missing comparisons and insufficient isolation of the feedback mechanism.

The evaluation omits several closely related methods discussed in Section 2, including CountDiffusion, CountCluster, Detection-Driven Object Count Optimization, and CountLoop. Moreover, comparisons with YOLO-Count on SDXL-Turbo and Counting Guidance on SD 1.4 involve different underlying generators, limiting conclusions about the relative effectiveness of the control methods themselves.

Although ATHENA-Static already evaluates a count-free control prompt, further ablations should isolate the effects of feedback-based prompt selection, the steering interval, and output normalization. In particular, fixed alternative-count prompts would help establish whether trajectory-specific count feedback is necessary. A G-DINO-based best-of-(N) or intermediate rejection-and-resampling baseline under a matched runtime budget is also important. These comparisons would clarify whether Feedback and Adaptive provide benefits beyond simple contrastive guidance and additional computation.

  • Practical relevance to newer generators and few-step sampling remains unclear.

The evaluation includes FLUX.2-dev and an unsteered Qwen-Image reference, but additional results from newer systems would better establish the remaining counting gap. Relevant examples include the open-weight Qwen-Image-2.1 and closed systems such as Nano Banana 2 or Nano Banana Pro and GPT Image 2.5, subject to their availability at submission. Although ATHENA requires access to intermediate model states and therefore cannot directly operate through closed image-generation APIs, unsteered results on ATHENA-COUNT would provide useful context, particularly for higher target counts and complex scenes.

More importantly, applicability to distilled few-step generators remains untested. For example, Z-Image-Turbo uses only eight model evaluations. ATHENA’s feedback-based variants rely on obtaining an informative intermediate reconstruction and restarting generation to intervene during an earlier, sufficiently malleable stage. The reported settings estimate counts after 20 or 30 steps in a 50-step trajectory. It remains unclear whether reliable count estimation and effective early intervention can coexist within a much shorter sampling schedule.

  • Incremental novelty beyond existing contrastive and feedback-based guidance.

Before norm matching, Equation (1) has the algebraic form of CFG with a modified text prompt replacing the unconditional branch. It closely follows the principle of contrastive-prompt guidance: Static removes the numerical specification, Feedback substitutes the detected count, and Adaptive adds a single heuristic adjustment of (\gamma). Intermediate count estimation followed by corrective guidance is also central to CountDiffusion.

The main methodological distinction therefore appears to be applying the correction to denoiser outputs rather than manipulating attention, which offers practical portability. However, demonstrating compatibility with several architectures does not establish whether this choice improves counting accuracy, image quality, or efficiency relative to attention-based correction. Comparisons with CountDiffusion and CountCluster on a shared backbone would help substantiate the contribution beyond implementation simplicity.

Questions for authors

See the Weaknesses for more details.

2 comments
AuthorsOpenReview ↗

Addressing reviewer ciiE concerns (part 1)

We thank the reviewer for their careful evaluation and constructive feedback. We appreciate their recognition of ATHENA's practical implementation, portability across different architectures, favorable accuracy–runtime trade-offs, and the value of our benchmark. We address the concerns below.

W1

Regarding the suggested baselines, complete implementations of CountDiffusion, CountCluster, and CountLoop were unavailable at the time of submission and rebuttal. CountDiffusion has no official public code, CountCluster's repository provides only a README, and CountLoop's repository states that its release is still under preparation, with key modules remaining placeholders. Our attempts to reproduce some of these methods yielded low accuracy, but without reference implementations, we could not determine whether this reflected their performance or implementation discrepancies. We therefore excluded these results to avoid unfair comparisons. Also, regarding Detection-Driven Object Count Optimization, we prioritized YOLO-Count as a related token-optimization method and CountGen, which already compares against Zafar et al.’s approach and reports higher counting accuracy.

Regarding backbone selection and comparison fairness, we agree that same-backbone comparisons are important. However, most existing baselines are not model-agnostic, a limitation ATHENA aims to address. We adapted Counting Guidance to FLUX.1-dev and SD 3.5 Large and evaluated ATHENA on SDXL to enable direct comparisons with Counting Guidance and CountGen, respectively. Moreover, YOLO-Count is included as a representative of gradient-based counting control in a single-step generation setting. In this method, iterative embedding optimization makes multi-step generation computationally impractical, so we retain its published SDXL-Turbo configuration, which has a runtime comparable to ATHENA on SDXL.

Regarding the Best-of-N comparison, following the reviewer’s suggestion, we evaluated detector-based Best-of-2 across all three backbones and datasets. This baseline generates two independent unsteered images and selects the one whose detected count is closest to the target. Complete results are provided below and in Appendix H.2 (Comparison with Best-of-2 Sampling), highlighted in blue. ATHENA-Adaptive achieves higher or comparable accuracy across all nine settings and lower wall-clock generation time in six. For example, on SD 3.5 Large and CoCoCount, ATHENA achieves 78.3% accuracy in 62.0 seconds, compared with 72.7% in 88.5 seconds for Best-of-2.

ModelMethodCoCoCount Acc. (%) ↑Time (s) ↓CoCoCount-E Acc. (%) ↑Time (s) ↓ATHENA-COUNT Acc. (%) ↑Time (s) ↓
FLUX.1-devBest-of-272.791.061.145.653.391.0
FLUX.1-devATHENA-Adaptive70.264.862.235.153.672.5
SD 3.5 LargeBest-of-272.788.562.748.353.187.9
SD 3.5 LargeATHENA-Adaptive78.362.065.638.156.170.4
SDXLBest-of-246.618.739.410.531.718.7
SDXLATHENA-Adaptive54.019.445.511.634.721.2

Beyond these results, ATHENA, like correction-based baselines, targets errors in the same seeded generation, aiming to preserve its scene structure and semantic content rather than replace it with an independently generated image. Moreover, additional feedback-guided correction attempts could further refine the steering strength γ\gamma. Resampling and steering are also complementary. Best-of-N can select among independently generated ATHENA candidates, combining ATHENA’s higher per-sample success rate with the benefits of candidate selection.

Regarding output normalization, we should clarify that it is essential for steering. Without it, changes in prediction magnitude can push subsequent latents outside the distribution expected by the denoiser, resulting in severe degradation or invalid images, which was the case in our preliminary experiments.

Regarding the steering interval, we tuned it using CoCoCount to balance counting accuracy and computational cost, then kept it fixed for each backbone across the other datasets. Although this does not replace a dedicated interval ablation, the consistent improvements suggest that the selected intervals transfer across datasets without additional tuning.

Regarding the effects of feedback-based prompt selection, we added an experiment on FLUX.1-dev and CoCoCount that retains feedback-based error detection but replaces the detected count in the control prompt with a random alternative count. Accuracy drops from 71.4% with ATHENA-Feedback to 65.8%, compared with 58.4% without steering. This directly supports using trajectory-specific feedback rather than an arbitrary count contrast.

AuthorsOpenReview ↗

Addressing reviewer ciiE concerns (part 2)

W2

Regarding newer generators, we thank the reviewer for raising this point. We evaluate four backbones, SDXL, SD 3.5 Large, FLUX.1-dev, and FLUX.2-dev, to cover different architectures and generation formulations, demonstrate transferability, and enable comparisons with model-specific baselines. FLUX.2-dev itself is recent, having been released in November 2025. The paper also includes the latest Qwen-Image available at the abstract deadline, and Qwen-Image-2.1 was released on September 20, two days later. Nevertheless, to address the reviewer’s concern, we have evaluated Qwen-Image-2.1 in the rebuttal phase, and the results show that it achieves 49.53% accuracy on CoCoCount and 36.11% on ATHENA-COUNT, confirming that counting remains challenging and ATHENA strategies on older backbones of FLUX.1-dev and SD 3.5 Large outperform this new generator.

Regarding proprietary models, we should note that these are black-box systems that may incorporate undisclosed processing beyond a standard T2I backbone. Without access to their mechanisms and computational requirements, we cannot make controlled comparisons. Their results can contextualize counting performance, but answer a different question from evaluating ATHENA.

Regarding few-step generators, we understand the reviewer’s concern. We argue that a shorter schedule still permits intermediate approximation; for example, an estimate can be obtained after four steps of an eight-step trajectory. The main challenge is effective steering within a compressed trajectory, which offers fewer opportunities for intervention. We acknowledge this limitation, but note that it presents a broader challenge for steering-based methods rather than one unique to ATHENA. Nevertheless, longer-step diffusion and flow models remain widely used for high-quality image generation, making them an important practical setting and the focus of our study.

W3

Regarding methodological novelty, we understand the reviewer’s concern about ATHENA’s CFG-style prediction contrast, but this equation is not our main contribution. Drawing on prior findings about T2I generation, we designed intermediate count estimation, feedback-informed control prompts, and selective early-stage steering as a unified framework that balances efficiency, effectiveness, and portability. Our evaluations show that this lightweight design outperforms more complex, model-specific approaches such as CountGen and computationally demanding methods such as Counting Guidance. Our methodological contribution is not simply constructing a count-specific prompt, but combining error detection, selective intervention, and adaptive correction into a unified framework.

Regarding comparisons, we kindly refer to our response to W1 for the implementation barriers affecting CountDiffusion and CountCluster. We have also extended Counting Guidance to FLUX.1-dev and SD 3.5 Large specifically to enable fair, same-backbone comparisons. Together with CountGen on SDXL, these comparisons demonstrate benefits in accuracy and efficiency beyond portability alone, although they do not replace direct comparisons with the unavailable methods.

Regarding evidence for the full design, Static, Feedback, and Adaptive form a progression that evaluates the benefit of increasingly informed correction. As reported in our response to W1, replacing the detected count with a random alternative in the control prompt yields 65.8% accuracy, compared with 71.4% for ATHENA-Feedback, on FLUX.1-dev and CoCoCount. This supports the benefit of trajectory-specific prompt construction. Appendix J.1 further reports comparable visual quality to unsteered generation, while evaluations across four backbones and four datasets show that the gains are not confined to one architecture or benchmark. Together, these results support the effectiveness of the complete framework beyond architectural portability alone.

We hope these additional experiments and clarifications address the reviewer’s concerns and better establish ATHENA’s contribution and practical value.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.