Weak Supervision Dynamic KL-Weighted Diffusion Models Guided by Large Language Models

Carter Scott; Frank Sanders; Julian Perry

REVIEW 5 major objections 5 minor 24 references

Weak Supervision Dynamic KL-Weighted Diffusion Models Guided by Large Language Models

T0 review · 5 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read LLM-guided diffusion with dynamic KL weighting beats GAN and prior diffusion baselines on COCO text-to-image generation.

desk verdict The core training loss in Eq. (6) is ill-posed, alpha_t is never defined, and the reported numbers are not credible; this paper should be desk rejected. read the letter →

arxiv 2502.00826 v1 pith:BJ7WZMFF submitted 2025-02-02 cs.CL

Julian Perry , Frank Sanders , Carter Scott This is my paper · ORCID

classification cs.CL

keywords text-to-imagegenerationdiffusionmodelslargelanguagedynamicKLweightingweak-to-strongguidancecross-attentionconditioningCOCOdatasethumanevaluation

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 text-to-image generation improves when a large language model supplies a fresh semantic embedding at every denoising step and the diffusion loss is weighted differently at each timestep. The authors report the best numbers among the methods they compare with on COCO: FID 30.5, Inception Score 5.4, and a 4.6 out of 5 average human rating, beating DDPM, CLIP-guided diffusion, AttnGAN, and T2I-DA. They also report about 20% lower training time and robustness to paraphrased prompts. If true, the result would make LLMs a practical control signal inside the diffusion process, not just a front-end that rewrites prompts.

What carries the argument

The machinery is the LLM-conditioned reverse diffusion step $p_\theta(x_{t-1}\mid x_t, z_t)$, where $z_t$ is a per-timestep text embedding computed by the language model and injected through a cross-attention map, together with the weighted variational objective $L_{\mathrm{weighted}} = \mathbb{E}_q\big[\sum_{t=1}^T \alpha_t D_{\mathrm{KL}}(q(x_t\mid x_{t-1}) \,\|\, p_\theta(x_{t-1}\mid x_t, z_t))\big]$. The time-dependent weight $\alpha_t$ is meant to emphasize coarse structure early and fine detail later; a momentum-based fine-tuning step on high-confidence self-generated samples is the remaining component. The paper defines $\alpha_t$ only symbolically and gives no schedule for it.

What would settle it

Run the same LLM-guided diffusion pipeline with $\alpha_t = 1$ for every timestep; if FID and IS stay at 30.5 and 5.4, the dynamic weighting is doing none of the work. Alternatively, any implementation that picks an explicit schedule and fails to reproduce Table I would show the published description is incomplete.

Watch

Extended reading notes

Core claim

The central claim is that a "Weak to Strong" training scheme, in which the LLM's influence is introduced progressively and the KL terms in the diffusion objective are reweighted by a time-dependent function, produces higher-quality, better-aligned images than existing text-to-image models. On the paper's reported numbers, the method has the lowest Fréchet Inception Distance (30.5) and the highest Inception Score (5.4) among the four baselines, and human raters prefer it by roughly a full point on a 1-5 scale. The ablation study attributes most of the gap to the LLM guidance and the dynamic KL weighting, and the authors additionally claim successful scaling to the OpenImages dataset. The paper does not provide the actual form of the weighting function.

Load-bearing premise

The entire reported improvement rests on the dynamic weighting function $\alpha_t$, but the paper never specifies its formula or schedule, so the method cannot be reproduced or checked as written.

Editorial extensions

If this is right

If the central claim holds, text-to-image systems can gain several FID points by conditioning every denoising step on an LLM embedding instead of treating the prompt as a one-shot input.
The reported 20% training-time reduction would make LLM-guided diffusion affordable for larger datasets and production deployments.
The reported generalization to novel prompts implies the LLM's semantic representation carries the model beyond memorized training captions.
The ablation ordering implies both the LLM guidance and the dynamic KL weighting are needed; removing either one degrades FID and IS by several points.

Reading between the lines

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

The sharpest place to test the paper is the missing schedule for $\alpha_t$: a re-run with any explicit schedule, compared with a constant-weight baseline, would separate the dynamic-weighting effect from the LLM conditioning effect.
The architecture as written is a cross-attention-conditioned diffusion model, so the distinguishing claim is really about the training curriculum; one could test whether ramping the LLM's influence with $t$ matters more than the embedding itself.
If the recipe transfers, per-timestep LLM conditioning should be tried in text-to-video and text-to-audio, where alignment must be maintained over time rather than in a single frame.

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.

Desk Editor's Note

read the letter

The stress-test note is right, and it is the whole ballgame. Equation (6) defines a KL divergence between q(x_t|x_{t-1}), a distribution over x_t, and p_theta(x_{t-1}|x_t,z_t), a distribution over x_{t-1}. Those live on different spaces; the divergence is not computable. No amount of specifying alpha_t can repair that, and since the entire method section rests on this loss, the method as written cannot be trained. This is a load-bearing flaw, not a gap in exposition.

What is the paper actually doing well? It names a real practical question: how to balance LLM conditioning against the diffusion model's own stochasticity over the course of generation. The weak-to-strong framing is a legitimate instinct, and the ablation study at least tries to isolate the contributions of LLM guidance and KL weighting. But that is where the credit ends. The central contribution, the dynamic weighting function alpha_t, is never given a formula or schedule. The promised 'alignment score' is never defined. There is no architecture detail, no hyperparameter list, no code, no data, no error bars.

The experimental section is worse than underspecified: it is internally suspicious. Inception Score around 4-5 on COCO is implausibly low; established text-to-image models score well above 20. The ablation row 'No LLM and No KL Weighting' is numerically identical to the DDPM baseline, which suggests the numbers were filled in rather than measured. The human evaluation reports an average rating of 4.6 for the proposed method versus 3.7 for the strongest baseline, with no participant demographics, no sample size per condition, no variance, no statistical test. The efficiency claims ('approximately 20% reduction in training time') and the scalability claims (OpenImages) appear out of thin air, with no supporting measurements.

The citation pattern is also sloppy: several references have nothing to do with the work (for example, an insect classification paper cited in the diffusion model overview), and the 'Weak to Strong' citation does not support the use made of it.

This paper is not ready for peer review. It has an invalid core objective, undefined key terms, and unverifiable empirical results. A serious referee would spend the entire time annotating fatal flaws. The right call is to desk reject. Nobody should build on this, and I would not cite it.

Referee Report

5 major / 5 minor

Summary. The paper proposes a text-to-image generation method that combines an LLM-conditioned diffusion model with a 'dynamic KL-weighting strategy.' The method is described as a weak-to-strong guidance approach in which an LLM provides per-timestep text embeddings z_t and a time-dependent weight α_t is applied to each KL term in the training loss. The authors report quantitative experiments on COCO (FID and IS) plus human evaluations and ablation studies, claiming that the method outperforms DDPM, CLIP-guided diffusion, AttnGAN, and T2I-DA. The paper also claims improved efficiency, robustness to textual variability, and scalability to other datasets and modalities.

Significance. If the proposed method were fully specified and validated, a robust LLM-guided diffusion objective with a principled dynamic KL weighting could be a useful contribution to controllable text-to-image generation. The paper also introduces an 'alignment score' that, if properly defined, could be valuable for measuring semantic coherence between text and generated images. However, the significance is entirely contingent on the correctness of the training objective and the credibility of the experimental evidence. Since the central loss is ill-posed, α_t is unspecified, and the experimental support is incomplete, the claimed improvements cannot currently be attributed to the proposed method. The paper does not ship code, does not provide derivations, and reports results without error bars or evaluation details, so there is no independent way to verify any of the surface-level claims.

major comments (5)

[Section III-C, Eq. (6)-(7)] The training objective is ill-posed. Equation (6) defines the loss as E_q[Σ_t D_KL(q(x_t|x_{t-1}) || p_θ(x_{t-1}|x_t, z_t))]. The left argument is a distribution over x_t conditioned on x_{t-1}, while the right argument is a distribution over x_{t-1} conditioned on x_t and z_t. These are distributions on different measurable spaces at different timesteps, so the KL divergence is not computable. The standard DDPM ELBO uses the tractable posterior q(x_{t-1}|x_t, x_0) on the left, not the forward transition q(x_t|x_{t-1}). Unless Eq. (6) is a typo, the described training procedure cannot be implemented, and every quantitative result in Tables I-V rests on an undefined objective.
[Section III-C, Eq. (7)] The dynamic weighting function α_t is never defined. The text states only that it is 'time-dependent' and 'designed to prioritize learning the early timesteps,' but no formula, schedule, or fitting procedure is given. Because α_t multiplies the ill-posed KL term in Eq. (7), no specification of α_t can fix the underlying problem; moreover, even if Eq. (6) were corrected, the choice of α_t would be an untested free parameter. The claimed benefit of the weighting strategy therefore cannot be evaluated.
[Table III] The ablation study contains a row, 'No LLM and No KL Weighting,' with FID 42.1 and IS 4.2, which is exactly identical to the DDPM baseline row in Table I. This suggests either that the ablation is not actually a removal of both components from the proposed model, or that the table entries were copied. In either case, the result undermines the claim that both LLM guidance and dynamic KL weighting are 'critical to the success of our method,' because the ablated model is indistinguishable from the standard DDPM baseline.
[Introduction, Section I, and abstract] The paper repeatedly promises a 'newly proposed alignment score' that 'quantifies the semantic coherence between text and image,' listing it as a key contribution and a motivation for the evaluation framework. However, this score is never defined, implemented, or reported anywhere in the experiments. The human evaluation tables (Tables II and V) ask participants to rate realism, relevance, and aesthetic quality, but these are not the promised quantitative alignment score. The absence of this central metric leaves the headline claim of improved text-image alignment without direct quantitative support.
[Section IV-B, Table I] The quantitative evaluation is not reproducible and the numbers are not contextualized. The paper does not report the image resolution, the number of generated samples, the specific COCO split (beyond '80-20'), the random seeds, or any error bars or confidence intervals for FID and IS. Without these details, the reported differences (e.g., FID 33.2 vs. 30.5) cannot be assessed for statistical significance. The manuscript also omits any comparison with modern text-to-image diffusion models such as Stable Diffusion or DALL-E, which limits the relevance of the claim that the method 'significantly outperforms existing approaches.'

minor comments (5)

[Abstract] The abstract contains grammatical errors, e.g., 'we presents a novel method' and 'a hybrid approach aimed at achieving both higher quality and efficiency in image synthesis from text descriptions' is a sentence fragment. The writing should be thoroughly revised.
[Section III-B, Eq. (5)] The cross-attention mechanism is described only as 'Attention(xt, zt)' with no formula, no definition of query/key/value projections, and no explanation of how the attention map at is used to modify the image features. This is insufficient for reproducing the method.
[Section II] Much of the related-work section is off-topic. References such as [5] (InsectMamba), [15] (Cedille), [16] (Goldfish), [20] (LLMs in bioinformatics), and [21] (psycholinguistics) have no clear connection to text-to-image generation or LLM-guided diffusion, and the discussion of them distracts from the paper's focus.
[Section IV-D and Table III] The ablation study lacks a description of what 'No KL Weighting Strategy' means: is α_t fixed to 1, removed entirely, or something else? Without this specification, the reader cannot interpret the difference between the 'Full Model' and 'No KL Weighting Strategy' rows.
[Section IV-F] The claim of 'approximately 20%' training-time reduction is presented with no timing measurements, no comparison of wall-clock time, no hardware configuration, and no convergence curves. The claim should either be removed or backed by actual measurements.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are unverifiable due to an undefined α_t and an ill-posed KL term, but no derivation reduces to its own inputs.

full rationale

The derivation chain does not reduce to its own inputs. The forward process in Eqs. (1)-(2) and the reverse process in Eqs. (3)-(4) are standard definitions, and the LLM conditioning in Eqs. (4)-(5) is a modeling choice rather than a definition of the reported FID, IS, or human-evaluation outcomes. The dynamic weight α_t in Eq. (7) is never specified, so the claimed gains cannot be verified from the equations, and Eq. (6) compares distributions over different timesteps and is ill-posed; these are correctness and reproducibility defects, not circular reductions. No parameter is fitted to a subset of the reported metrics and then presented as a prediction of that same subset. The references, including the "Weak to Strong" citation [1], are not authored by the present authors and are not used to forbid alternative methods. The unspecified alignment score is also a completeness gap rather than a renaming of the target result. Under the hard rule that circularity must be exhibited by a specific reduction, no such reduction can be quoted, so the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The framework depends on the standard DDPM assumptions, an unproven use of time-dependent text embeddings, and an entirely unspecified alpha_t function that carries all the claimed gains. No parameters are disclosed, and the only 'new' entity, the alignment score, is never defined.

free parameters (3)

alpha_t (dynamic KL weighting) = unspecified
The central contribution is alpha_t in Eq. 7, but no functional form, schedule, or fitting procedure is described; all performance claims depend on it.
High-confidence sample threshold = unspecified
The momentum-based fine-tuning step uses high-confidence samples, but the paper does not define the confidence measure or threshold.
Model and training hyperparameters = unspecified
Diffusion backbone, LLM choice, optimizer, batch size, learning rate, and noise schedule are never reported.

assumptions (3)

standard math Standard DDPM forward and reverse processes (Eqs. 1-3) are valid.
The paper restates the canonical denoising diffusion formulation without proof.
domain assumption Time-dependent LLM embeddings z_t can guide each reverse step meaningfully.
The paper assumes text embeddings vary across timesteps and improve control, but offers no evidence or mechanism.
ad hoc to paper An unspecified alpha_t improves training.
The method's entire novelty rests on alpha_t, which is never defined; this is a load-bearing assumption.

invented entities (1)

Alignment score
purpose: Claimed as a new metric to quantify semantic consistency between text and image.
Mentioned in the introduction and contributions, but never defined or reported in the experiments, so it has no independent verification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Weak Supervision Dynamic KL-Weighted Diffusion Models Guided by Large Language Models." pith.science (2026). https://pith.science/paper/BJ7WZMFF

@misc{pith2026250200826,
  author       = {Pith},
  title        = {Pith review of: Weak Supervision Dynamic KL-Weighted Diffusion Models Guided by Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BJ7WZMFF}},
  note         = {Machine review of arXiv:2502.00826}
}

read the original abstract

In this paper, we presents a novel method for improving text-to-image generation by combining Large Language Models (LLMs) with diffusion models, a hybrid approach aimed at achieving both higher quality and efficiency in image synthesis from text descriptions. Our approach introduces a new dynamic KL-weighting strategy to optimize the diffusion process, along with incorporating semantic understanding from pre-trained LLMs to guide the generation process. The proposed method significantly improves both the visual quality and alignment of generated images with text descriptions, addressing challenges such as computational inefficiency, instability in training, and robustness to textual variability. We evaluate our method on the COCO dataset and demonstrate its superior performance over traditional GAN-based models, both quantitatively and qualitatively. Extensive experiments, including ablation studies and human evaluations, confirm that our method outperforms existing approaches in terms of image realism, relevance to the input text, and overall aesthetic quality. Our approach also shows promise in scalability to other multimodal tasks, making it a versatile solution for a wide range of generative applications.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 12 canonical work pages

[1]

Weak to strong generaliza tion for large language models with multi-capabilities,

Y . Zhou, J. Shen, and Y . Cheng, “Weak to strong generaliza tion for large language models with multi-capabilities,” in The Thirteenth International Conference on Learning Representations , 2024. [Online]. Available: https://openreview.net/pdf?id=N1vYivuSKq

work page 2024
[2]

Master ing text-to-image diffusion: Recaptioning, planning, and gen erating with multimodal llms,

L. Y ang, Z. Y u, C. Meng, M. Xu, S. Ermon, and C. Bin, “Master ing text-to-image diffusion: Recaptioning, planning, and gen erating with multimodal llms,” in F orty-ﬁrst International Conference on Machine Learning, 2024

work page 2024
[3]

Dil ightnet: Fine-grained lighting control for diffusion-based image g eneration,

C. Zeng, Y . Dong, P . Peers, Y . Kong, H. Wu, and X. Tong, “Dil ightnet: Fine-grained lighting control for diffusion-based image g eneration,” in ACM SIGGRAPH 2024 Conference Papers , 2024, pp. 1–12

work page 2024
[4]

Adding conditional co ntrol to text-to-image diffusion models,

L. Zhang, A. Rao, and M. Agrawala, “Adding conditional co ntrol to text-to-image diffusion models,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 3836–3847

work page 2023
[5]

Insectmamba: Insec t pest classiﬁcation with state space model,

Q. Wang, C. Wang, Z. Lai, and Y . Zhou, “Insectmamba: Insec t pest classiﬁcation with state space model,” arXiv preprint arXiv:2404.03611 , 2024

arXiv 2024
[6]

An overview of diffusion models: Applications, guided generation, stati stical rates and optimization,

M. Chen, S. Mei, J. Fan, and M. Wang, “An overview of diffusion models: Applications, guided generation, stati stical rates and optimization,” CoRR, vol. abs/2404.07771, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2404.07771

arXiv doi:10.48550/arxiv.2404.07771 2024
[7]

Diffusio n model with representation alignment for protein inverse folding ,

C. Wang, Y . Zhou, Z. Zhai, J. Shen, and K. Zhang, “Diffusio n model with representation alignment for protein inverse folding ,” arXiv preprint arXiv:2412.09380, 2024

arXiv 2024
[8]

Where to dif fuse, how to diffuse, and how to get back: Automated learning for multivariate diffusions,

R. Singhal, M. Goldstein, and R. Ranganath, “Where to dif fuse, how to diffuse, and how to get back: Automated learning for multivariate diffusions,” in The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023. [Online]. Available: https://openreview.net/forum?id=osei3IzUia

work page 2023

Show all 24 references

[9]

Diffusion models fo r non- autoregressive text generation: A survey,

Y . Li, K. Zhou, W. X. Zhao, and J. Wen, “Diffusion models fo r non- autoregressive text generation: A survey,” in Proceedings of the Thirty- Second International Joint Conference on Artiﬁcial Intell igence, IJCAI 2023, 19th-25th August 2023, Macao, SAR, China . ijcai.org, 202...

2023
[10]

Diffusion on dynamic contact networks w ith indirect transmission links,

M. Shahzamal, “Diffusion on dynamic contact networks w ith indirect transmission links,” CoRR, vol. abs/1906.02856, 2019. [Online]. Available: http://arxiv.org/abs/1906.02856

1906 arXiv
[11]

Diffusion mo dels are evolutionary algorithms,

Y . Zhang, B. Hartl, H. Hazan, and M. Levin, “Diffusion mo dels are evolutionary algorithms,” CoRR, vol. abs/2410.02543, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2410.02543

2024 doi arXiv
[12]

Claret: Pre-training a correlation-aware context-to-event transformer for eve nt-centric gener- ation and classiﬁcation,

Y . Zhou, T. Shen, X. Geng, G. Long, and D. Jiang, “Claret: Pre-training a correlation-aware context-to-event transformer for eve nt-centric gener- ation and classiﬁcation,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: ...

2022
[13]

Fine-grained distillation for long document ret rieval,

Y . Zhou, T. Shen, X. Geng, C. Tao, J. Shen, G. Long, C. Xu, a nd D. Jiang, “Fine-grained distillation for long document ret rieval,” in Proceedings of the AAAI Conference on Artiﬁcial Intelligen ce, vol. 38, no. 17, 2024, pp. 19 732–19 740

2024
[14]

Eventbe rt: A pre- trained model for event correlation reasoning,

Y . Zhou, X. Geng, T. Shen, G. Long, and D. Jiang, “Eventbe rt: A pre- trained model for event correlation reasoning,” in Proceedings of the ACM W eb Conference 2022, 2022, pp. 850–859

2022
[15]

Cedille: A large autoregres sive french language model,

M. M¨ uller and F. Laurent, “Cedille: A large autoregres sive french language model,” CoRR, vol. abs/2202.03371, 2022. [Online]. Available: https://arxiv.org/abs/2202.03371

2022 arXiv
[16]

Goldﬁsh: Monolingual language models for 350 languages,

T. A. Chang, C. Arnett, Z. Tu, and B. K. Bergen, “Goldﬁsh: Monolingual language models for 350 languages,” CoRR, vol. abs/2408.10441, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2408.10441

2024 doi arXiv
[17]

Less is more: Vision representation compression for efﬁcient video gene ration with large language models,

Y . Zhou, J. Zhang, G. Chen, J. Shen, and Y . Cheng, “Less is more: Vision representation compression for efﬁcient video gene ration with large language models,” 2024

2024
[18]

Style-aware contrastive learning for multi-style image captioning,

Y . Zhou and G. Long, “Style-aware contrastive learning for multi-style image captioning,” in Findings of the Association for Computational Linguistics: EACL 2023 , 2023, pp. 2257–2267

2023
[19]

Modeling event-pair relations in external knowledge graphs for scri pt reasoning,

Y . Zhou, X. Geng, T. Shen, J. Pei, W. Zhang, and D. Jiang, “ Modeling event-pair relations in external knowledge graphs for scri pt reasoning,” Findings of the Association for Computational Linguistics : ACL-IJCNLP 2021, 2021

2021
[20]

Large language models in bioinformatics: applications an d perspectives,

J. Liu, M. Y ang, Y . Y u, H. Xu, K. Li, and X. Zhou, “Large language models in bioinformatics: applications an d perspectives,” CoRR, vol. abs/2401.04155, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2401.04155

2024 doi arXiv
[21]

Beyond t he limitations of any imaginable mechanism: large language mo dels and psycholinguistics,

C. J. Houghton, N. Kazanina, and P . Sukumaran, “Beyond t he limitations of any imaginable mechanism: large language mo dels and psycholinguistics,” CoRR, vol. abs/2303.00077, 2023. [Online]. Available: https://doi.org/10.48550/arXiv.2303.00077

2023 doi arXiv
[22]

Improv ing zero-shot cross-lingual transfer for multilingual questi on answering over knowledge graph,

Y . Zhou, X. Geng, T. Shen, W. Zhang, and D. Jiang, “Improv ing zero-shot cross-lingual transfer for multilingual questi on answering over knowledge graph,” in Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Ling uistics: H...

2021
[23]

How good are commercial large language models on african languages?

J. Ojo and K. Ogueji, “How good are commercial large language models on african languages?” in Proceedings of the 4th W orkshop on African Natural Language Processing, AfricaNL P@ICLR 2023, Kigali, Rwanda, May 1, 2023 , 2023. [Online]. Available: https://openreview.net/pdf?id=...

2023
[24]

Lost in translation: Large l anguage models in non-english content analysis,

G. Nicholas and A. Bhatia, “Lost in translation: Large l anguage models in non-english content analysis,” CoRR, vol. abs/2306.07377, 2023. [Online]. Available: https://doi.org/10.48550/arXiv.2306.07377 6

2023 doi arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Weak to strong generaliza tion for large language models with multi-capabilities,

Y . Zhou, J. Shen, and Y . Cheng, “Weak to strong generaliza tion for large language models with multi-capabilities,” in The Thirteenth International Conference on Learning Representations , 2024. [Online]. Available: https://openreview.net/pdf?id=N1vYivuSKq

work page 2024

[2] [2]

Master ing text-to-image diffusion: Recaptioning, planning, and gen erating with multimodal llms,

L. Y ang, Z. Y u, C. Meng, M. Xu, S. Ermon, and C. Bin, “Master ing text-to-image diffusion: Recaptioning, planning, and gen erating with multimodal llms,” in F orty-ﬁrst International Conference on Machine Learning, 2024

work page 2024

[3] [3]

Dil ightnet: Fine-grained lighting control for diffusion-based image g eneration,

C. Zeng, Y . Dong, P . Peers, Y . Kong, H. Wu, and X. Tong, “Dil ightnet: Fine-grained lighting control for diffusion-based image g eneration,” in ACM SIGGRAPH 2024 Conference Papers , 2024, pp. 1–12

work page 2024

[4] [4]

Adding conditional co ntrol to text-to-image diffusion models,

L. Zhang, A. Rao, and M. Agrawala, “Adding conditional co ntrol to text-to-image diffusion models,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 3836–3847

work page 2023

[5] [5]

Insectmamba: Insec t pest classiﬁcation with state space model,

Q. Wang, C. Wang, Z. Lai, and Y . Zhou, “Insectmamba: Insec t pest classiﬁcation with state space model,” arXiv preprint arXiv:2404.03611 , 2024

arXiv 2024

[6] [6]

An overview of diffusion models: Applications, guided generation, stati stical rates and optimization,

M. Chen, S. Mei, J. Fan, and M. Wang, “An overview of diffusion models: Applications, guided generation, stati stical rates and optimization,” CoRR, vol. abs/2404.07771, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2404.07771

arXiv doi:10.48550/arxiv.2404.07771 2024

[7] [7]

Diffusio n model with representation alignment for protein inverse folding ,

C. Wang, Y . Zhou, Z. Zhai, J. Shen, and K. Zhang, “Diffusio n model with representation alignment for protein inverse folding ,” arXiv preprint arXiv:2412.09380, 2024

arXiv 2024

[8] [8]

Where to dif fuse, how to diffuse, and how to get back: Automated learning for multivariate diffusions,

R. Singhal, M. Goldstein, and R. Ranganath, “Where to dif fuse, how to diffuse, and how to get back: Automated learning for multivariate diffusions,” in The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023. [Online]. Available: https://openreview.net/forum?id=osei3IzUia

work page 2023