Pith. sign in

REVIEW 5 major objections 4 minor 64 references

Rectify Then Diffuse: Disentangling Concepts Before Denoising Trajectory Unfolds

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Rectify-then-Diffuse claims that compositional failures in text-to-image models are an early spatial-allocation problem, fixable with one bounded gradient step on the initial noise before denoising begins.

desk verdict New initial-latent rectification method with strong empirical gains, but a missing random-displacement control leaves its mechanism claim under-supported. read the letter →

arxiv 2608.03135 v1 pith:MBFGSQOZ submitted 2026-08-04 cs.CV cs.AI

Ning Zhu , An Chen , Mengfei Zhao , Juntao Xu , Jingze Liang , Boyuan Gu , Liang-Jian Deng This is my paper
classification cs.CVcs.AI
keywords text-to-imagediffusioncompositionalgenerationcross-attentionanalysisinitiallatentrectificationsoftoverlapdisentanglementtraining-freeinferenceconceptbindingsampling
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

Text-to-image diffusion models can render one concept well but often omit or merge concepts when several are requested. This paper claims the roots of that failure are visible before denoising starts: a single pilot forward pass at a near-pure-noise timestep shows each concept's cross-attention map, and concepts whose maps overlap heavily at this stage tend to stay coupled through denoising and end up missing or fused. RTD therefore treats composition as a boundary-condition problem—it applies one bounded, normalized gradient update to the initial latent, maximizing a soft intersection-over-union separation objective, and then lets the original sampler run untouched. If correct, compositional fidelity can be improved with no training, no per-step guidance, and no layout input, at roughly 6.3% extra inference cost. The paper reports state-of-the-art BLIP-VQA, ImageReward, and human-evaluation scores on AE-Bench, T2I-CompBench, and RareBench, including a 45.8% BLIP-VQA improvement over CO3 on object pairs with 2.3x faster runtime.

What carries the argument

The load-bearing object is the pilot concept attention map: cross-attention probabilities from one frozen U-Net pass at the first DDIM timestep ($t_{\mathrm{pilot}} = 980$), averaged over heads and $32\times32$ modules and summed over each concept's tokens. SOD turns these maps into soft occupancy maps $M^{(k)}$ and scores every pair with soft IoU, $O_{ij} = \langle M^{(i)}, M^{(j)}\rangle / (\lVert M^{(i)}\rVert_1 + \lVert M^{(j)}\rVert_1 - \langle M^{(i)}, M^{(j)}\rangle + \epsilon)$, defining separation $S(x_T) = 1 - \text{average}\, O_{ij}$. IGR then applies the single update $x'_T = x_T + \rho \lVert x_T\rVert_2\, g / \max(\lVert g\rVert_2, \epsilon)$ with $g = \nabla_{x_T} S$, so the c

What would settle it

Take a large batch of multi-concept prompts, compute pilot-map soft IoU at t=980 across seeds, and compare with final-image concept presence; if high-overlap seeds are not the ones that fail, or if a same-norm random bounded perturbation of the initial latent reproduces RTD's BLIP-VQA gains, the overlap signal is not the actionable bottleneck the paper claims.

Watch

Extended reading notes

Core claim

The paper's central claim is that multi-concept generation fails through an early spatial-allocation bottleneck: before any image structure exists, prompt-conditioned cross-attention already assigns each concept a provisional spatial support, and concepts whose supports strongly overlap at this stage remain coupled through denoising, producing omission, fusion, or swapped attributes. RTD measures this coupling with Soft-Overlap Disentanglement (SOD), which max-min normalizes each concept's pilot attention map into a soft occupancy map and computes pairwise soft intersection-over-union; maximizing one minus the average overlap gives a layout-agnostic, differentiable separation signal. Isotrop

Load-bearing premise

That the concept attention maps from one diagnostic pass at a near-pure-noise timestep faithfully predict how concepts will be spatially allocated across the whole denoising trajectory, so that reducing their overlap actually changes the final image.

Editorial extensions

If this is right

  • A single extra forward-backward pass at initialization replaces per-step attention guidance or score correction: RTD adds 6.3% overhead over vanilla SDXL and runs 2.3x faster than CO3.
  • Because the rectified latent is sampler-agnostic, the same one-step correction improves DDIM and DPM++ 2M at 10, 20, and 50 steps, and transfers to SD 2.1 without retuning.
  • Early-overlap (S-IoU at step 5) becomes a cheap diagnostic that tracks compositional fidelity: RTD reports the lowest early overlap and the best composition scores on all three benchmarks.
  • The method imposes no explicit layout, so object scale, arrangement, and interaction remain controlled by the pretrained generator; separation is a soft prior on initial spatial support, not a hard box or mask.

Reading between the lines

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

  • Editorial inference: the same pilot-overlap signal could be used as a pre-sampling failure predictor—compute soft IoU at $t_{\mathrm{pilot}}$ for many seeds and flag high-overlap candidates before spending inference budget on them.
  • Testable extension left implicit in the paper: conditioning the separation objective on relation semantics (holding, wearing, occlusion) would let RTD respect interactions that require spatial overlap, where generic separation may be counterproductive.
  • Editorial inference: a norm-matched control—perturbing the initial latent by the same $\rho\lVert x_T\rVert_2$ in a random direction—would directly test whether the overlap gradient, rather than any bounded latent shift, is the active ingredient behind the reported gains.
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 / 4 minor

Summary. The paper proposes Rectify-then-Diffuse (RTD), a training-free method that intervenes only on the initial latent of a text-to-image diffusion model. RTD first runs a diagnostic pilot forward pass at a high-noise timestep, extracts per-concept cross-attention maps, and defines a Soft-Overlap Disentanglement (SOD) objective that penalizes pairwise soft IoU between normalized concept maps. It then applies one Isotropic Gradient Rectification (IGR) update — a bounded, gradient-normalized displacement of the initial latent — and proceeds with standard DDIM sampling. Experiments on AE-Bench, T2I-CompBench, and RareBench report improved BLIP-VQA, ImageReward, and human-evaluation scores over CO3 and other baselines, with roughly 6.3% inference overhead. The paper argues that early concept allocation is an actionable bottleneck and that correcting it before denoising improves compositional fidelity.

Significance. If the mechanism is correct, RTD is conceptually appealing: it is training-free, sampler-agnostic, requires only one extra forward/backward pass, and reframes compositional failure as an initial-condition problem. The method is well specified, the ablation of SOD versus IGR is informative, and the cross-solver, cross-step, and cross-backbone experiments are a strength. However, the central causal claim — that reducing pilot-map overlap via the SOD gradient direction is what drives the gains — is not yet supported by the evidence: there is no control for generic initial-latent perturbation, the reported S-IoU5 metric is the same soft-IoU that SOD optimizes, and the headline CO3/SDXL numbers are borrowed from another paper. The A-O subset also shows consistent regressions relative to CO3, which is not reflected in the abstract's unqualified state-of-the-art claim. These issues are fixable with additional experiments and more careful reporting.

major comments (5)
  1. [§3.3, Eq. (5)] The causal claim requires a random-displacement control. The IGR update x'_T = x_T + ρ||x_T||₂ g/||g||₂ is a bounded perturbation of a near-pure-noise latent. Diffusion sampling is sensitive to initial-latent changes, so the gains in Tables 1–2 could arise from any sufficiently large perturbation rather than from the semantic content of the SOD gradient. Please add a control with the same ρ and same norm but a random (or fixed, content-free) unit direction, and ideally a per-prompt comparison between the change in S-IoU5 and the change in final metrics. This is load-bearing: without it, the paper's mechanism is not distinguished from generic initial-condition sensitivity.
  2. [§4.2, Table 1 footnote] The abstract's headline numbers — 45.8% BLIP-VQA and 19.6% ImageReward improvement over CO3 on AE-Bench O-O — depend on CO3 and SDXL values taken directly from Dutta et al. (2026). Because RTD is evaluated with three fixed seeds, SDXL at 1024×1024, and 50-step DDIM under this paper's protocol, the comparison may not be controlled. Please re-run all main baselines under the identical evaluation pipeline, or clearly mark which numbers are not recomputed and qualify the reported gains accordingly.
  3. [§4.2, S-IoU5 definition] The paper reports 'Early Overlap S-IoU5' as evidence for the mechanism, but this metric is the same soft intersection-over-union that SOD optimizes in Eq. (3). Reporting lower S-IoU5 for RTD is therefore a manipulation check, not an independent outcome. To support the early-allocation bottleneck, provide an independent measure of early spatial separation (e.g., IoU on thresholded maps at a later step, or final segmentation masks) and, more importantly, a per-prompt correlation between pilot overlap at t=980 and final compositional success/failure.
  4. [§4.2, Table 1] The A-O subset results contradict the unqualified 'state-of-the-art' claim. On animal–object prompts, RTD is below CO3 on both BLIP-VQA (0.8422 vs 0.8878) and ImageReward (1.5881 vs 1.6743). The paper says it achieves SOTA in 'four of six' settings, but the abstract and conclusion state overall SOTA without this caveat. Please either discuss why A-O does not improve or soften the claim; this is also relevant to the paper's hypothesis, since A-O is a large fraction of AE-Bench.
  5. [§4.1 and all result tables] All results are means over three seeds with no error bars, per-prompt variance, or significance tests. Given that the reported gains over CO3 are sometimes small (e.g., RareBench Complex ImageReward 1.2476 vs 1.2211), the absence of statistical support makes it hard to assess whether the improvements are robust. Add seed-level standard deviations or per-prompt paired tests, particularly for the headline comparisons.
minor comments (4)
  1. [§4.2, first sentence] 'we further report the mean pairwise Step-5 Soft IoU computed from concept attention maps as SOD' should say 'as S' or 'using the soft-IoU defined in Eq. (3)', since SOD is the name of the objective, not the metric variable.
  2. [Table 4 title] The table title carries a superscript footnote marker '1', but no corresponding footnote is provided in the text. Either add the note or remove the marker.
  3. [§5 and Appendix F] The main-paper Limitations section is only one sentence; important limitations in Appendix F — phrase-internal modifiers, relation-dependent overlap, dependence on the pretrained model, and architectural scope — should be summarized in the main paper to give readers an accurate scope.
  4. [§4.4, Figure 5] ρ=0.02 and t_pilot=980 are selected based on the best aggregate performance on the same test benchmarks. Please describe the selection protocol (e.g., separate validation split or a small tuning set) or acknowledge that the sensitivity analysis was conducted on the evaluation benchmarks themselves.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild self-definitional diagnostic (S-IoU5 restates the SOD objective); central fidelity claims rest on external metrics.

  1. self definitional [Section 4.2, 'To quantify early concept competition...' and Table 1 caption; Eq. (3)-(4)]
    "To quantify early concept competition, we further report the mean pairwise Step-5 Soft IoU computed from concept attention maps as SOD. ... RTD's consistently lowest S-IoU5 indicates early feature-level concept separation, which is positively associated with its stronger multi-concept generation scores."

    S-IoU5 is computed with the same soft-IoU formula that SOD's separation objective S is designed to maximize (Eq. 3-4). RTD's update is gradient ascent on that objective at t_pilot=980, so a reduced S-IoU5 at Step 5 is a manipulation check guaranteed in the direction the method was built to move, not an independent confirmation of the mechanism. The paper phrases this as 'indicates early feature-level concept separation' and links it to its success, which is by-construction consistency. The main fidelity claims, however, are backed by external BLIP-VQA, ImageReward, and human evaluation, so this self-definitional element is not the load-bearing evidence for the central result.

full rationale

The core RTD derivation is self-contained: pilot cross-attention maps A^(k) are functions of x_T; SOD defines a differentiable soft-IoU separation objective S(x_T); IGR computes the normalized gradient and applies one bounded update x'_T = x_T + ρ||x_T||_2 g/||g||_2. This is a standard first-order constrained optimization step, and the subsequent evaluation uses external benchmarks (AE-Bench, T2I-CompBench, RareBench) with BLIP-VQA, ImageReward, and human ratings, none of which are constructed from the SOD objective. The paper's self-citations (e.g., Su et al. 2025, Zhu et al. 2025a/b, Ma et al. 2025) appear only in motivational or background contexts and are not load-bearing for the method's claims. The only noticeable circular element is the S-IoU5 diagnostic, which is literally the overlap quantity SOD minimizes; reporting it as evidence of separation is a manipulation check rather than an independent validation. That is a mild self-definitional element, but it does not make the central claim circular because the headline improvements are measured with external, non-fitted metrics. Hyperparameter choices (rho=0.02, t_pilot=980) are selected using the AE-Bench benchmark itself, which is a benchmark-selection caveat rather than a circularity of the derivation chain.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method rests on the empirical link between pilot attention overlap and final image fidelity, plus standard differentiability of the diffusion U-Net. The two fitted hyperparameters (rho, t_pilot) are chosen on the evaluation benchmark. No new entities are postulated.

free parameters (3)
  • Relative correction ratio rho = 0.02
    Controls the L2 displacement of the initial latent relative to its norm; selected from sensitivity analysis on AE-Bench (Fig. 5a).
  • Pilot timestep t_pilot = 980
    First timestep in the 50-step DDIM schedule; selected from sensitivity analysis (Fig. 5b).
  • Pilot attention spatial resolution = 32x32
    Only cross-attention modules with 32x32 spatial grid are recorded; chosen to avoid mixing map granularities (Sec. B.2).
assumptions (5)
  • domain assumption Cross-attention maps at high-noise timestep reflect the spatial allocation of concepts that persists through denoising.
    Invoked in Sec. 1 and 3.1 to justify using pilot maps as the rectification signal.
  • domain assumption Reducing pairwise soft overlap between normalized pilot concept maps improves compositional fidelity of the final image.
    Central empirical premise behind the SOD objective (Sec. 3.2).
  • standard math The frozen denoising U-Net is differentiable with respect to the initial latent, so the gradient of S can be computed.
    Required for g = grad_{x_T} S in Sec. 3.3.
  • domain assumption A single first-order gradient step with step size rho times the latent norm is adequate to change concept allocation.
    IGR solves the linearized constrained problem (Eq. B.2-B.4); the paper assumes the linear approximation holds at rho=0.02.
  • domain assumption BLIP-VQA and ImageReward are valid proxies for compositional fidelity.
    Used as primary evaluation metrics following CO3 (Sec. 4.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rectify Then Diffuse: Disentangling Concepts Before Denoising Trajectory Unfolds." pith.science (2026). https://pith.science/paper/MBFGSQOZ

@misc{pith2026260803135,
  author       = {Pith},
  title        = {Pith review of: Rectify Then Diffuse: Disentangling Concepts Before Denoising Trajectory Unfolds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MBFGSQOZ}},
  note         = {Machine review of arXiv:2608.03135}
}
abstract

Text-to-image diffusion models can generate individual concepts well, but they often omit or merge concepts incorrectly with multiple concepts. We trace these failures to an early coordination bottleneck: before denoising begins, prompt-conditioned attention may allocate different concepts to strongly overlapping spatial support, which can keep their attention coupled as denoising proceeds. This observation motivates treating compositional generation as a boundary-condition problem rather than repeatedly controlling the evolving trajectory. To this end, we propose Rectify-then-Diffuse (RTD), a training-free framework that rectifies the initial allocation once before standard denoising. Firstly, we propose Soft-Overlap Disentanglement (SOD), which converts normalized overlap between pilot concept maps into a differentiable and layout-agnostic separation objective. Secondly, we introduce Isotropic Gradient Rectification (IGR), which normalizes the SOD gradient and applies a bounded latent displacement with a consistent scale across prompts and initializations. Extensive experiments show that RTD achieves state-of-the-art compositional fidelity and robust gains. On the AE-Bench object pair subset, RTD improves BLIP-VQA by 45.8% and ImageReward by 19.6% over CO3 while running 2.3$\times$ faster. Code will be released at https://github.com/Z-yiwei/rectify-then-diffuse

Figures

Figures reproduced from arXiv: 2608.03135 by the authors.

Figure 1
Figure 1. Qualitative comparison on AE-Bench. From top to bottom: SDXL, CO3 and RTD. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of RTD. RTD extracts pilot concept maps, uses SOD to measure overlap, and applies IGR update. The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison on T2I-CompBench. RTD [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Concept attention evolution during denoising. Rows compare CO3 and RTD. Colored borders match the highlighted [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Sensitivity to (a) relative correction ratio [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 46 canonical work pages

  1. [1]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Denoising Diffusion Probabilistic Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  2. [2]

    International Conference on Learning Representations (ICLR) , year=

    Denoising Diffusion Implicit Models , author=. International Conference on Learning Representations (ICLR) , year=

  3. [3]

    High-Order Flow Matching: Unified Framework and Sharp Statistical Rates , volume =

    Su, Maojiang and Hu, Jerry Yao-Chieh and Lee, Yi-Chen and Zhu, Ning and Chung, Jui-Hui and Wu, Shang and Song, Zhao and Chen, Minshuo and Liu, Han , booktitle =. High-Order Flow Matching: Unified Framework and Sharp Statistical Rates , volume =

  4. [4]

    Proceedings of the 32nd International Conference on Machine Learning , pages=

    Deep Unsupervised Learning using Nonequilibrium Thermodynamics , author=. Proceedings of the 32nd International Conference on Machine Learning , pages=

    work page

  5. [5]

    Diffusion Models Beat

    Dhariwal, Prafulla and Nichol, Alexander , booktitle=. Diffusion Models Beat

  6. [6]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    High-Resolution Image Synthesis with Latent Diffusion Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  7. [7]

    International Conference on Learning Representations (ICLR) , year=

    Podell, Dustin and English, Zion and Lacey, Kyle and Blattmann, Andreas and Dockhorn, Tim and M. International Conference on Learning Representations (ICLR) , year=

    work page

  8. [8]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    CSAL-3D: Cold-Start Active Learning for 3D Medical Image Segmentation via SSL-Driven Uncertainty-Reinforced Diversity Sampling , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    work page

Show all 64 references
  1. [9]

    Hierarchical Text-Conditional Image Generation with

    Aditya Ramesh and Prafulla Dhariwal and Alex Nichol and Casey Chu and Mark Chen , journal=. Hierarchical Text-Conditional Image Generation with

  2. [10]

    Technical Report, OpenAI , year=

    Improving Image Generation with Better Captions , author=. Technical Report, OpenAI , year=

  3. [11]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  4. [12]

    Unsupervised Anomaly Detection for Aircraft PRSOV With Random Projection-Based Inner Product Prediction , year=

    Peng, Dandan and Zhu, Ning and Han, Te and Chen, Zhuyun and Liu, Chenyu , journal=. Unsupervised Anomaly Detection for Aircraft PRSOV With Random Projection-Based Inner Product Prediction , year=

  5. [13]

    Proceedings of the 41st International Conference on Machine Learning , volume =

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis , author=. Proceedings of the 41st International Conference on Machine Learning , volume =

  6. [14]

    Proceedings of the 38th International Conference on Machine Learning , pages=

    Learning Transferable Visual Models from Natural Language Supervision , author=. Proceedings of the 38th International Conference on Machine Learning , pages=

  7. [15]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Attention Is All You Need , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  8. [16]

    Li, Junnan and Li, Dongxu and Savarese, Silvio and Hoi, Steven , booktitle=

  9. [17]

    ACM Transactions on Graphics , volume=

    Attend-and-Excite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models , author=. ACM Transactions on Graphics , volume=. doi:10.1145/3592116 , year=

    doi

  10. [18]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Linguistic Binding in Diffusion Models: Enhancing Attribute Correspondence through Attention Map Alignment , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  11. [19]

    Agarwal, Aishwarya and Karanam, Srikrishna and Joseph, K J and Saxena, Apoorv and Goswami, Rohan and Srinivasan, Balaji Vasan , booktitle=

  12. [20]

    34th British Machine Vision Conference 2023,

    Divide & Bind Your Attention for Improved Generative Semantic Nursing , author=. 34th British Machine Vision Conference 2023,

    2023

  13. [21]

    2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    Training-Free Layout Control with Cross-Attention Guidance , author=. 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    2024

  14. [22]

    An Adversarial Training Framework Based on Unsupervised Feature Reconstruction Constraints for Crystalline Silicon Solar Cells Anomaly Detection , year=

    Zhu, Ning and Wang, Jing and Zhang, Ying and Wang, Huan and Han, Te , journal=. An Adversarial Training Framework Based on Unsupervised Feature Reconstruction Constraints for Crystalline Silicon Solar Cells Anomaly Detection , year=

  15. [23]

    European Conference on Computer Vision (ECCV) , pages=

    Compositional Visual Generation with Composable Diffusion Models , author=. European Conference on Computer Vision (ECCV) , pages=

  16. [24]

    The Eleventh International Conference on Learning Representations , year=

    Training-Free Structured Diffusion Guidance for Compositional Text-to-Image Synthesis , author=. The Eleventh International Conference on Learning Representations , year=

  17. [25]

    International Conference on Learning Representations (ICLR) , year=

    Prompt-to-Prompt Image Editing with Cross Attention Control , author=. International Conference on Learning Representations (ICLR) , year=

  18. [26]

    2025 , author =

    Deep adaptive wavelet autoencoder with mutually independent empirical cumulative distribution for unsupervised motor anomaly detection , journal =. 2025 , author =

    2025

  19. [27]

    Kwon, Gihyun and Ye, Jong Chul , booktitle =

  20. [28]

    International Conference on Learning Representations (ICLR) , year=

    Steer Away From Mode Collisions: Improving Composition In Diffusion Models , author=. International Conference on Learning Representations (ICLR) , year=

  21. [29]

    Guo, Xiefan and Liu, Jinlin and Cui, Miaomiao and Li, Jiankai and Yang, Hongyu and Huang, Di , booktitle=

  22. [30]

    What the

    Tang, Raphael and Liu, Linqing and Pandey, Akshat and Jiang, Zhiying and Yang, Gefei and Kumar, Karun and Stenetorp, Pontus and Lin, Jimmy and Ture, Ferhan , booktitle=. What the

  23. [31]

    Adversarial Frequency Component Reconstruction Constraint for Helicopter Vibration Signal Anomaly Detection: An Unsupervised Dual-Domain Approach , year=

    Zhu, Ning and Dong, Tianzhi Xu and Peng, Dandan , journal=. Adversarial Frequency Component Reconstruction Constraint for Helicopter Vibration Signal Anomaly Detection: An Unsupervised Dual-Domain Approach , year=

  24. [32]

    Multi-Concept

    Tunanyan, Hazarapet and Xu, Dejia and Navasardyan, Shant and Zhang, Zhifei and Shi, Humphrey , journal=. Multi-Concept

  25. [33]

    Huang, Kaiyi and Sun, Kaiyue and Xie, Enze and Li, Zhenguo and Liu, Xihui , booktitle=

  26. [34]

    Xu, Jiazheng and Liu, Xiao and Wu, Yuchen and Tong, Yuxuan and Li, Qinkai and Ding, Ming and Tang, Jie and Dong, Yuxiao , booktitle=

  27. [35]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    SUGFW: A SAM-Based Uncertainty-Guided Feature Weighting Framework for Cold Start Active Learning , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

  28. [36]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-Free Diffusion Guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    arXiv

  29. [37]

    Journal of the American Statistical Association , volume=

    Tweedie's Formula and Selection Bias , author=. Journal of the American Statistical Association , volume=

  30. [38]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations , pages=

    Stanza: A Python Natural Language Processing Toolkit for Many Human Languages , author=. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations , pages=

  31. [39]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Token Merging for Training-Free Semantic Binding in Text-to-Image Synthesis , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  32. [40]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Magnet: We Never Know How Text-to-Image Diffusion Models Work, Until We Learn How Vision-Language Models Function , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  33. [41]

    arXiv preprint arXiv:2508.03441 , year=

    MedCAL-Bench: A Comprehensive Benchmark on Cold-Start Active Learning with Foundation Models for Medical Image Analysis , author=. arXiv preprint arXiv:2508.03441 , year=

  34. [42]

    International Conference on Learning Representations (ICLR) , pages =

    Rare-to-Frequent: Unlocking Compositional Generation Power of Diffusion Models on Rare Concepts with LLM Guidance , author=. International Conference on Learning Representations (ICLR) , pages =

  35. [43]

    2025 , pages=

    Lu, Cheng and Zhou, Yuhao and Bao, Fan and Chen, Jianfei and Li, Chongxuan and Zhu, Jun , journal=. 2025 , pages=

    2025

  36. [44]

    Nichol, Alexander Quinn and Dhariwal, Prafulla and Ramesh, Aditya and Shyam, Pranav and Mishkin, Pamela and McGrew, Bob and Sutskever, Ilya and Chen, Mark , booktitle=

  37. [45]

    Chen, Junsong and Yu, Jincheng and Ge, Chongjian and Yao, Lewei and Xie, Enze and Wang, Zhongdao and Kwok, James and Luo, Ping and Lu, Huchuan and Li, Zhenguo , booktitle=

  38. [46]

    Salih Meral, Tuna Han and Simsar, Enis and Tombari, Federico and Yanardag, Pinar , booktitle=

  39. [47]

    2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    Dense Text-to-Image Generation with Attention Modulation , author=. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    2023

  40. [48]

    Xie, Jinheng and Li, Yuexiang and Huang, Yawen and Liu, Haozhe and Zhang, Wentian and Zheng, Yefeng and Shou, Mike Zheng , booktitle=

  41. [49]

    2023 , volume=

    Lian, Long and Li, Boyi and Yala, Adam and Darrell, Trevor , journal=. 2023 , volume=

    2023

  42. [50]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Diffusion Self-Guidance for Controllable Image Generation , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  43. [51]

    2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    Improving Sample Quality of Diffusion Models Using Self-Attention Guidance , author=. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    2023

  44. [52]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Visual Programming for Step-by-Step Text-to-Image Generation and Evaluation , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  45. [53]

    , booktitle=

    Hu, Yushi and Liu, Benlin and Kasai, Jungo and Wang, Yizhong and Ostendorf, Mari and Krishna, Ranjay and Smith, Noah A. , booktitle=

  46. [54]

    Ghosh, Dhruba and Hajishirzi, Hannaneh and Schmidt, Ludwig , booktitle=

  47. [55]

    Wu, Xindi and Yu, Dingli and Huang, Yangsibo and Russakovsky, Olga and Arora, Sanjeev , booktitle=

  48. [56]

    Zhou, Dewei and Li, You and Ma, Fan and Zhang, Xiaoting and Yang, Yi , booktitle=

  49. [57]

    Li, Mingcheng and Hou, Xiaolu and Liu, Ziyang and Yang, Dingkang and Qian, Ziyun and Chen, Jiawei and Wei, Jinjie and Jiang, Yue and Xu, Qingyao and Zhang, Lihua , booktitle=

  50. [58]

    Dat, Do Huu and Hyeon-Woo, Nam and Mao, Po-Yuan and Oh, Tae-Hyun , booktitle=

  51. [59]

    AAAI Conference on Artificial Intelligence , year=

    Compositional Text-to-Image Synthesis with Attention Map Control of Diffusion Models , author=. AAAI Conference on Artificial Intelligence , year=

  52. [60]

    2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Attention Calibration for Disentangled Text-to-Image Personalization , author=. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    2024

  53. [61]

    Shentu, Junjie and Watson, Matthew and Al Moubayed, Noura , journal=

  54. [62]

    Lim, Habin and Won, Yeongseob and Seo, Juwon and Park, Gyeong-Moon , booktitle=

  55. [63]

    Unified Compositional Controller: A Training-Free Framework for Highly Controllable Text-to-Image Generation , author=. Inf. Sci. , volume=. doi:10.1016/j.ins.2026.123380 , year=

    2026

  56. [64]

    arXiv preprint arXiv:2406.07844 , year=

    Understanding and Mitigating Compositional Issues in Text-to-Image Generative Models , author=. arXiv preprint arXiv:2406.07844 , year=

    arXiv

Pith tools

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