Pith. sign in

REVIEW 5 major objections 6 minor 27 references

OptiPrune: Boosting Prompt-Image Consistency with Attention-Guided Noise and Dynamic Token Selection

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that a training-free combination of attention-guided noise optimization and similarity-based token pruning improves prompt-image consistency and efficiency in text-to-image diffusion models.

desk verdict Plausible training-free framework, but the headline claims outrun the evidence: the efficiency claim has no measurements, and the alignment gains over InitNO are 0.1 points without statistics. read the letter →

arxiv 2507.00789 v1 pith:UT7D7AIH submitted 2025-07-01 cs.CV

classification cs.CV
keywords EfficientDiffusionModelsInitialNoiseOptimizationTokenPruningprompt-imageconsistencyattentionmapstraining-freeCLIPsimilaritytext-to-image
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that semantic misalignment and computational cost in text-to-image diffusion models can be addressed together rather than as separate trade-offs. It proposes OptiPrune, a training-free framework with two components: LatentMapper, which optimizes the initial noise toward 'valid' regions defined by attention-map diagnostics, and SimPrune, which prunes redundant tokens in self-attention layers and recovers them by copying from a small set of base tokens. The authors claim this joint design achieves state-of-the-art prompt-image consistency on benchmarks such as Animal-Animal, with margins around 0.1 percentage points over the strongest noise-optimization baseline, and that the pruning accelerates both noise optimization and generation. If the claim holds, diffusion models could be made both more semantically faithful and cheaper to run without retraining.

What carries the argument

The load-bearing machinery is the pair of latent-validity scores in Eqs. (3)–(4), which turn the qualitative goals 'subjects are not neglected' and 'subjects are not entangled' into differentiable optimization targets, together with the patch-wise base-token selection and maximum-similarity recovery rule of SimPrune. The scores drive the inner-loop gradient updates in LatentMapper; the patch-based selection enforces spatial coverage of retained tokens, and the recovery by feature copying makes pruning cheap and approximately reversible. The KL term in the joint loss keeps the optimized noise near the Gaussian prior, which the paper argues preserves trainability and compatibility with standard samplers.

What would settle it

A concrete test would be a human-rating study on generated images whose noise passes versus fails the validity thresholds, or a sweep of the thresholds $\tau_c$ and $\tau_s$ in the validity check while watching whether CLIP similarity and human judgments track the threshold; if images from 'invalid' noise rate as well as or better than 'valid' noise, the central alignment claim is falsified. A second, direct falsifier for the efficiency claim is a wall-clock latency comparison on the same GPU with and without SimPrune, since the paper's tables report no runtimes.

Watch

Extended reading notes

Core claim

The central discovery is that the same attention information that reveals which latent regions are semantically problematic can also identify which tokens are redundant, so noise optimization and token pruning reinforce each other. LatentMapper partitions latent space into valid and invalid noise using a cross-attention response score $S_{\text{CrossAttn}}$, which measures whether every subject in the prompt receives enough activation, and a self-attention conflict score $S_{\text{SelfAttn}}$, which measures spatial overlap between subjects. It optimizes a Gaussian noise distribution under a KL constraint so the optimized noise stays close to the standard prior. SimPrune then selects spatially distributed base tokens per image patch, prunes tokens most similar to those bases, injects Gaussian noise into the similarity scores for diversity, and recovers pruned tokens by copying output features of their nearest base token. The paper's ablations attribute subject mixing to removal of LatentMapper and attribute binding failures to removal of SimPrune, supporting the claim that the two components work synergistically.

Load-bearing premise

The framework assumes that the two attention-based scores faithfully measure prompt-image alignment, so pushing those scores below thresholds produces images that humans or CLIP will judge as better aligned; the thresholds themselves are never specified or ablated.

Editorial extensions

If this is right

  • If OptiPrune works as claimed, text-to-image models can get better prompt-image alignment without any fine-tuning or per-model training, because both modules are training-free and data-agnostic.
  • The pruning is confined to self-attention layers, so the framework preserves cross-attention text-image alignment while still reducing the main computational bottleneck of attention.
  • The same valid-noise criterion could be reused as a cheap early-exit check during sampling: stop optimizing as soon as both scores drop below threshold.
  • Because the optimized noise stays near a Gaussian prior, the framework is likely to plug into existing DDPM/DDIM schedulers and latent diffusion variants without special accommodation.
  • On resource-constrained hardware, the claimed speedups from token pruning could make real-time text-to-image generation feasible, which is the deployment scenario the paper motivates.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported margins over the strongest baseline are about 0.1 percentage points on several CLIP metrics, so a seed-variance analysis would show how stable those margins are across prompts and seeds.
  • A natural extension is to make the validity thresholds adaptive or learned rather than fixed, since the paper leaves $\tau_c$ and $\tau_s$ unspecified.
  • The self-attention conflict score could be repurposed as a training reward or guidance signal beyond inference-time filtering, since it appears to capture spatial entanglement directly.
  • The efficiency claim is testable with a direct latency benchmark; if SimPrune's overhead in selecting base tokens and copying features offsets its savings in self-attention, the speed advantage could be smaller than implied.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. This paper presents OptiPrune, a training-free framework for text-to-image diffusion that couples (1) LatentMapper, an attention-guided optimization of the initial latent noise that repartitions the latent space based on cross-attention response and self-attention conflict scores, with (2) SimPrune, a cosine-similarity token pruning method with Gaussian noise injection and feature-copy recovery. The authors evaluate on Animal-Animal, Animal-Object, and Object-Object datasets using CLIP full-prompt, minimum-object, and text-text similarity, reporting small gains over InitNO (about 0.1 percentage points) and claiming 'significantly reduced computational cost.' The paper does not report any runtime, FLOP, latency, or memory measurements, and the conclusion concedes that the noise-optimization stage incurs non-negligible overhead.

Significance. If fully validated, the idea of jointly optimizing initial noise and pruning tokens is a useful conceptual contribution to efficient and aligned text-to-image generation, and the training-free design is appealing. The paper's strength is its concise formulation of latent-space diagnostics and a token-selection heuristic that are both simple and plausible. However, the current evidence is insufficient: the consistency gain over InitNO is tiny and lacks statistical support, the efficiency claim is unsupported by any measurement, and the validity criteria for the optimized noise are essentially the same scores being minimized. With the required measurements and ablations, the work could become a solid empirical contribution; as it stands, the central two-part claim is not established.

major comments (5)
  1. [Abstract, Sec. 5, Tables 1-3] The abstract and conclusion claim 'significantly reduced computational cost,' but no wall-clock time, FLOP count, GPU memory, throughput, or latency measurement appears anywhere in the manuscript. The conclusion (Sec. 5) states that the initial noise optimization 'still incurs non-negligible computational overhead,' and LatentMapper's inner/outer optimization loops (Sec. 3.2) require repeated UNet forward and backward passes. The efficiency half of the central claim is therefore unsupported; provide runtime and FLOP comparisons against InitNO and SiTo, or remove the efficiency claim.
  2. [Sec. 4.3, Tables 1-3] The reported improvements over InitNO are 0.1 percentage points on Full Prompt and Minimum Object Similarity in most rows (e.g., Animal-Animal 33.5 vs 33.4; Animal-Object 36.2 vs 36.1; Object-Object 36.4 vs 36.3). No error bars, variance, or significance tests are reported for any of the 16-image-averaged metrics. These differences are likely within run-to-run noise and do not by themselves establish state-of-the-art prompt-image consistency. Report confidence intervals, standard deviations, or statistical tests across multiple seeds.
  3. [Sec. 3.2, Eqs. (3)-(6)] LatentMapper directly minimizes SCrossAttn and SSelfAttn, which are exactly the criteria used to classify noise as 'valid' (SCrossAttn < τc and SSelfAttn < τs). The in-proxy improvement is therefore by construction, and the thresholds τc and τs are never specified or ablated. To support the claim that this produces semantically better images, validate the diagnostics against independent measures (e.g., human evaluation or the CLIP metrics used in Sec. 4.3) and report sensitivity to τc, τs, and λ.
  4. [Sec. 3.3, Eqs. (7)-(10)] SimPrune requires an O(N²) pairwise cosine similarity matrix and a per-patch argmax over all tokens before pruning can begin. The paper does not analyze whether this overhead is amortized across layers or timesteps, nor does it report any speedup measurement for SimPrune itself. Provide a complexity analysis and measured runtime/FLOPs for SimPrune alone and for the full OptiPrune pipeline, or the claimed acceleration from token pruning remains unquantified.
  5. [Sec. 4.3 and Sec. 5] The conclusion claims the method 'outperforming methods like InitNO and SiTo in CLIP similarity metrics,' but SiTo is absent from all quantitative tables (Tables 1-3). Include SiTo and other recent token-pruning baselines in the quantitative comparisons, or restrict the performance claim to the baselines actually evaluated.
minor comments (6)
  1. [Sec. 4.3 and Sec. 4.4] Sections 4.3 and 4.4 share the same title 'Quantitative Evaluation' and contain nearly identical text; the duplication suggests an incomplete revision and should be resolved by merging the sections.
  2. [Sec. 3.3] The heading 'T oken Puring' appears to be a typo for 'Token Pruning.'
  3. [Eq. (10)] The notation 'argtopK' is not defined; please describe the selection procedure in words or provide a formal definition.
  4. [Sec. 3.2, Eq. (4)] The notation (xi, yi) for spatial centroids reuses the symbol y, which is also used for the prompt tokens yi; please switch to different coordinate symbols such as (ux, uy) to avoid confusion.
  5. [Sec. 4.3] The text says 'outperforming InitNO by 0.1%' and 'a 0.7% lead' when describing differences between similarity percentages; these are percentage points, not percent, and should be phrased accordingly.
  6. [Fig. 3] The legend 'sd v1 v2 ours' is ambiguous; spell out 'Stable Diffusion,' 'V1 (no LatentMapper),' 'V2 (no SimPrune),' and 'full OptiPrune.' Also, the axis label 'animals_objects' is inconsistent with 'Animal-Object' used in the text.

Circularity Check

1 steps flagged · score 2.0 of 10

LatentMapper's 'valid noise' is valid by its own optimization objective, but the headline CLIP-benchmark claim is externally evaluated; no load-bearing circularity.

  1. self definitional [Sec. 3.2, Eqs. (3)-(6)]
    "Noise is valid when both SCrossAttn < τc and SSelfAttn < τs hold simultaneously. ... Ljoint = SCrossAttn + SSelfAttn + λ · KL (N (µ, Σ) ∥ N(0, I)) ... The first two terms minimize subject neglect and mixing, while the Kullback-Leibler divergence constrains the optimized distribution to remain proximate to the original Gaussian prior."

    The validity predicate is defined by thresholds on SCrossAttn and SSelfAttn (Eqs. 3-4), and the optimization objective Ljoint (Eq. 6) is exactly the sum of these two scores plus a KL term. Thus any converged output is 'valid' by construction: the method's own success criterion is the training loss. Claiming that LatentMapper 'mitigates subject neglect and feature entanglement' based on these scores is a restatement of the objective, not independent evidence. However, the paper's headline quality claim is checked on external CLIP image-text/text-text benchmarks (Tables 1-3), so this self-definitional component does not force the central result. The Gaussian-prior preservation is likewise a direct consequence of adding the KL term, not a measured property.

full rationale

The central prompt-image consistency claim is supported by external CLIP-based benchmarks on Animal-Animal, Animal-Object, and Object-Object datasets, which are independent of the attention diagnostics used inside LatentMapper. The only notable circularity is internal: noise is declared 'valid' by the same SCrossAttn/SSelfAttn scores that the optimization minimizes, so the in-proxy improvement is by construction. This does not propagate to the final evaluation, so it is mild rather than load-bearing. There are no self-citations, no imported uniqueness theorems, and no fitted parameter renamed as a prediction. The efficiency claim ('significantly reduced computational cost') lacks any runtime, FLOP, or latency measurement and is even qualified in the conclusion by 'non-negligible computational overhead,' but that is an evidentiary gap, not circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim carries six hand-set or unreported hyperparameters and five domain assumptions about attention diagnostics, copy-based recovery, noise injection, prior preservation, and the evaluation metric. No new physical entities are introduced.

free parameters (6)
  • tau_c (subject neglect threshold) = not reported
    Threshold in Eq. (3) dividing valid from invalid noise; never specified or ablated in the paper.
  • tau_s (self-attention conflict threshold) = not reported
    Threshold in Eq. (4) for spatial entanglement; never specified or ablated.
  • lambda (KL weight) = not reported
    Weight of KL term in Eq. (6); controls how close the optimized noise stays to N(0,1); no value or sensitivity analysis given.
  • sigma (noise injection scale) = not reported
    Std of Gaussian noise added to similarity scores in Eq. (8); no value reported, affecting pruning selection and reproducibility.
  • s (patch size) = not reported
    Patch size for base token selection in Eq. (9); hyperparameter with no stated value.
  • pruning ratio K (gamma) = gamma = 0.4
    Pruning ratio used in Sec. 4.2 comparisons; no tuning procedure or sensitivity reported.
assumptions (5)
  • domain assumption Cross-attention response and self-attention overlap capture subject neglect and spatial entanglement.
    Sec. 3.2 states these are 'geometrically interpretable metrics'; the optimization directly minimizes them and the paper assumes this improves alignment.
  • domain assumption Tokens adjacent to selected base tokens exhibit minimal recovery error.
    Sec. 3.3 states this to justify the copy-based recovery; no experiments validate the claim in isolation.
  • ad hoc to paper Injected Gaussian noise to similarity scores improves generalization.
    Eq. (8) adds noise 'to ensure diverse selection'; the paper provides no evidence or theory for this design.
  • domain assumption The KL constraint keeps the optimized noise within the generative prior without hurting quality.
    Eq. (6) uses KL to preserve the Gaussian prior, but the paper does not measure the distribution shift of generated images.
  • domain assumption CLIP similarity is a faithful measure of prompt-image consistency.
    CLIP-based metrics used throughout Sec. 4; standard in the field but an assumption about what 'consistency' means.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OptiPrune: Boosting Prompt-Image Consistency with Attention-Guided Noise and Dynamic Token Selection." pith.science (2026). https://pith.science/paper/UT7D7AIH

@misc{pith2026250700789,
  author       = {Pith},
  title        = {Pith review of: OptiPrune: Boosting Prompt-Image Consistency with Attention-Guided Noise and Dynamic Token Selection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UT7D7AIH}},
  note         = {Machine review of arXiv:2507.00789}
}
read the original abstract

Text-to-image diffusion models often struggle to achieve accurate semantic alignment between generated images and text prompts while maintaining efficiency for deployment on resource-constrained hardware. Existing approaches either incur substantial computational overhead through noise optimization or compromise semantic fidelity by aggressively pruning tokens. In this work, we propose OptiPrune, a unified framework that combines distribution-aware initial noise optimization with similarity-based token pruning to address both challenges simultaneously. Specifically, (1) we introduce a distribution-aware noise optimization module guided by attention scores to steer the initial latent noise toward semantically meaningful regions, mitigating issues such as subject neglect and feature entanglement; (2) we design a hardware-efficient token pruning strategy that selects representative base tokens via patch-wise similarity, injects randomness to enhance generalization, and recovers pruned tokens using maximum similarity copying before attention operations. Our method preserves the Gaussian prior during noise optimization and enables efficient inference without sacrificing alignment quality. Experiments on benchmark datasets, including Animal-Animal, demonstrate that OptiPrune achieves state-of-the-art prompt-image consistency with significantly reduced computational cost.

Figures

Figures reproduced from arXiv: 2507.00789 by the authors.

Figure 1
Figure 1. Illustration of our joint framework integrating LatentMapper and SimPrune. (a) LatentMapper partitions latent space using cross-attention and self-attention map to optimize initial noise toward valid regions. (b) SimPrune prunes redundant tokens in self-attention layers through spatially distributed base token selection and similarity￾based recovery. The optimized noise and pruned token set are jointly leveraged for… view at source ↗
Figure 2
Figure 2. Visualization of ablation study.For all methodologies, images are synthesized using same text prompts and the same seed. Our approach demonstrates strong corre￾spondence with text prompts, all the while preserving high visual realism in generated outputs. images using BLIP [23] and compute the similarity between these captions and the input prompt. The quantitative results in Tables 1, 2, and 3 comprehensively demon… view at source ↗
Figure 3
Figure 3. Average CLIP Image-Text Similarity, including Full Prompt Similarity and Minimum Object Similarity, and Average CLIP Text-Text Similarity are reported for the quantitative measurement. Higher is better. purple balloon and the white clock are clearly separated in space. Each object has its own distinct boundaries, and there is no overlapping or merging of the two objects. The purple color of the balloon does not blee… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 21 canonical work pages

  1. [1]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10684–10695 (2022)

  2. [2]

    https://arxiv.org/abs/2311.15127

    Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A.: Stable video diffusion: Scaling latent video diffusion models to large datasets(2023). https://arxiv.org/abs/2311.15127

  3. [3]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Guo, X., Liu, J., Cui, M., Li, J., Yang, H., Huang, D.: Initno: Boosting text-to-image diffusion models via initial noise optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 9380–9389 (2024)

  4. [4]

    Advances in Neural Information Processing Systems 37, 125487–125519 (2024)

    Eyring, L., Karthik, S., Roth, K., Dosovitskiy, A., Akata, Z.: Reno: Enhancing one- step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems 37, 125487–125519 (2024)

  5. [5]

    https://arxiv.org/abs/2405.14828

    Xu, K., Zhang, L., Shi, J.: Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models(2024). https://arxiv.org/abs/2405.14828

  6. [6]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025)

    Shen, F., Du, X., Gao, Y., Yu, J., Cao, Y., Lei, X., Tang, J.: IMAGHarmony: Con- trollable image editing with consistent object quantity and layout. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025)

  7. [7]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Bolya, D., Hoffman, J.: Token merging for fast stable diffusion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 4599– 4603 (2023)

  8. [8]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp

    Kim, M., Gao, S., Hsu, Y.-C., Shen, Y., Jin, H.: Token fusion: Bridging the gap be- tween token pruning and token merging. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 1383–1392 (2024)

Show all 27 references
  1. [9]

    K., Liu, Y.: Attention-driven training-free efficiency enhancement of diffusion models

    Wang, H., Liu, D., Kang, Y., Li, Y., Lin, Z., Jha, N. K., Liu, Y.: Attention-driven training-free efficiency enhancement of diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16080– 16089 (2024)

  2. [10]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, pp

    Zhang, E., Tang, J., Ning, X., Zhang, L.: Training-free and hardware-friendly ac- celeration for diffusion models via similarity-based token pruning. In: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 9878–9886 (2025)

  3. [11]

    Advances in neural information processing systems 33, 6840–6851 (2020) Abbreviated paper title 13

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840–6851 (2020) Abbreviated paper title 13

  4. [12]

    https://arxiv.org/abs/2010.02502

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models(2020). https://arxiv.org/abs/2010.02502

  5. [13]

    Advances in neural information processing systems 34, 11287–11302 (2021)

    Vahdat, A., Kreis, K., Kautz, J.: Score-based generative modeling in latent space. Advances in neural information processing systems 34, 11287–11302 (2021)

  6. [14]

    In: Advances in Neural Information Processing Systems, vol

    Shen, F., Tang, J.: ImagPose: A unified conditional framework for pose-guided person generation. In: Advances in Neural Information Processing Systems, vol. 37, pp. 6246–6266 (2024)

  7. [15]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025)

    Shen, F., Yu, J., Wang, C., Jiang, X., Du, X., Tang, J.: IMAGGarment-1: Fine- grained garment generation for controllable fashion design. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025)

  8. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025)

    Shen, F., Wang, C., Gao, J., Guo, Q., Dang, J., Tang, J., Chua, T.-S.: Long-term talking face generation via motion-prior conditional diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025)

  9. [17]

    https://arxiv.org/abs/2210.09461

    Bolya, D., Fu, C.-Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster(2022). https://arxiv.org/abs/2210.09461

  10. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Jin, P., Takanobu, R., Zhang, W., Cao, X., Yuan, L.: Chat-univi: Unified visual representation empowers large language models with image and video understand- ing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13700–13710 (2024)

  11. [19]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Song, E., Chai, W., Wang, G., Zhang, Y., Zhou, H., Wu, F., Chi, H., Guo, X., Ye, T., Zhang, Y., and others: Moviechat: From dense token to sparse memory for long video understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 182...

  12. [20]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Shen, F., Jiang, X., He, X., Ye, H., Wang, C., Du, X., Li, Z., Tang, J.: IMAGDressing-V1: Customizable virtual dressing. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 7, pp. 6795–6804 (2025)

  13. [21]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Shen, F., Ye, H., Liu, S., Zhang, J., Wang, C., Han, X., Yang, W.: Boosting con- sistency in story visualization with rich-contextual conditional diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 7, pp. 6785–6794 (2025)

  14. [22]

    ACM Trans- actions on Graphics (TOG) 42(4), 1–10 (2023)

    Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Trans- actions on Graphics (TOG) 42(4), 1–10 (2023)

  15. [23]

    In: International Conference on Machine Learning, pp

    Li, Junnan, Li, Dongxu, Xiong, Caiming, Hoi, Steven: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and gener- ation. In: International Conference on Machine Learning, pp. 12888–12900. PMLR, (2022)

  16. [24]

    B.: Compositional visual gen- eration with composable diffusion models

    Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J. B.: Compositional visual gen- eration with composable diffusion models. In: European Conference on Computer Vision, pp. 423–439. Springer (2022)

  17. [25]

    E., Wang, W

    Feng, W., He, X., Fu, T., Jampani, V., Akula, A., Narayana, P., Basu, S., Wang, X. E., Wang, W. Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis(2022).https://arxiv.org/abs/2212.05032

  18. [26]

    Li, Y., Keuper, M., Zhang, D., Khoreva, A.: Divide & bind your attention for improved generative semantic nursing(2023).https://arxiv.org/abs/2307.10864

  19. [27]

    V.: A-star: Test-time attention segregation and retention for text-to-image synthesis

    Agarwal, A., Karanam, S., Joseph, K., Saxena, A., Goswami, K., Srinivasan, B. V.: A-star: Test-time attention segregation and retention for text-to-image synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 2283–2293 (2023)

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.