Cross-Cultural Fashion Design via Interactive Large Language Models and Diffusion Models

Amina Grant; Jeffrey Lee; Spencer Ramsey

REVIEW 5 major objections 4 minor 31 references

Cross-Cultural Fashion Design via Interactive Large Language Models and Diffusion Models

T0 review · 5 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Adding an LLM prompt-refinement stage and a weak-supervision filter to a latent diffusion model produces higher-quality, more culturally diverse fashion images than StyleGAN2, DALL-E 2, or the base LDM, with FID 7.80 on expanded…

desk verdict An under-specified, incremental LLM-plus-diffusion pipeline whose headline numbers don't match its own tables; the evaluation is too shaky to support the claims. read the letter →

arxiv 2501.15571 v1 pith:Q5SZTIJC submitted 2025-01-26 cs.CL

Spencer Ramsey , Amina Grant , Jeffrey Lee This is my paper · ORCID

classification cs.CL

keywords fashionimagegenerationlatentdiffusionmodelslargelanguagepromptrefinementweaksupervisionfilteringculturaldiversitytext-to-imagealignmentDeepFashion+

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 a generative pipeline which lets a large language model enrich user prompts, a latent diffusion model render the images, and a weak-supervision filter admit only useful noisy labels can produce fashion images that are both higher quality and more culturally diverse than existing generators. On an expanded DeepFashion+ dataset the pipeline reports a Frechet Inception Distance of 7.80 and an Inception Score of 34.02, beating StyleGAN2, DALL-E 2, and the unmodified latent diffusion model on both metrics and on human ratings of quality, cultural diversity, and prompt relevance. The point of the claim is practical: culturally inclusive fashion generation is currently limited by biased datasets, scarce full annotations, and weak text-image alignment, and the paper argues these can be addressed together without expensive manual labeling.

What carries the argument

The load-bearing mechanism is a three-stage pipeline. A pre-trained large language model rewrites and expands each user prompt with cultural and stylistic detail; a latent diffusion model, consisting of a variational autoencoder (VAE) that maps images into a compressed latent space, a UNet denoiser that reverses added noise, and cross-attention that injects the refined prompt, generates the image; and a weak-supervision filtering module scores weakly labeled samples by the cosine similarity between a text encoder's reading of the prompt and an image encoder's reading of the candidate image, keeping only samples above a threshold $\tau$. The training objective combines the standard denoising loss with a prompt-consistency loss and a reconstruction loss. The filtering module is what makes noisy or weakly labeled data usable, and the prompt-consistency loss is what keeps the generated image aligned with the LLM-refined description.

What would settle it

Run the same comparison on a publicly released version of the expanded DeepFashion+ dataset with StyleGAN2, DALL-E 2, and the base LDM fine-tuned under identical data and prompt settings. If the full model's FID does not reproduce near 7.80, or if ablations do not show the reported degradation without prompt refinement (9.45) and without weak-supervision filtering (10.12), the central claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that the two added components are what drive the improvement. The authors report that removing the LLM prompt-refinement stage raises FID from 7.80 to 9.45, removing the weak-supervision filter raises it to 10.12, and removing both raises it to 11.30, while Inception Score falls correspondingly from 34.02 to 32.10, 31.45, and 30.70. Human raters give the full pipeline 4.7/5 for quality, 4.5/5 for cultural diversity, and 4.6/5 for relevance, and a CLIP-based alignment score reaches 0.91 against 0.82 for the base LDM. The paper interprets these results as showing that LLM-guided semantic enrichment and selective use of weakly labeled data together make diffusion-based fashion generation scalable and more inclusive.

Load-bearing premise

The load-bearing premise is that the 'expanded DeepFashion+ dataset' enriched with culturally diverse styles is a concrete, available artifact and that all baselines were fine-tuned on the same data with the same prompt conditions, so the reported FID and IS gaps come from the proposed modules rather than from different training setups.

Editorial extensions

If this is right

Fashion image generation can be improved substantially without full manual annotations, because the weak-supervision filter lets noisy labels contribute to training.
A designer can feed a short prompt such as 'African print dress' and receive images with richer cultural detail, because LLM refinement expands the prompt before generation.
The need for fully captioned datasets is reduced, lowering the cost of scaling fashion generation to more styles and regions.
The ablations show that both the prompt-refinement and filtering modules are individually necessary for the reported performance.

Reading between the lines

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

A testable extension is to sweep the filtering threshold $\tau$; the paper fixes one threshold, and an optimal operating point would show how much weakly labeled data can be tolerated before quality drops.
The paper's stated limitation with rare cultural contexts suggests the bottleneck may be dataset coverage rather than model architecture; measuring FID separately for held-out underrepresented styles would test this.
If the reported gains replicate under controlled baselines, the same LLM-refinement-plus-weak-filter recipe could apply to other weakly labeled text-to-image domains such as architecture or product design, though the paper does not claim this.

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

Short version: this is an under-specified, incremental re-packaging of LLM-guided diffusion with a CLIP-threshold filter, and the experiments don't stand up. The strongest thing I can say for it is that the writing is organized and it openly states limitations.

What is actually new: not much. The method is straightforward—LLM refines prompts, LDM generates, and a cosine-similarity filter selects weakly labeled samples. The equations are standard. The paper does cite the earlier interactive fashion LLM+LDM work [17] and does not pretend the combination is unprecedented. The ablation study at least attempts to isolate components, and the limitations paragraph is honest about reliance on large pretrained models and struggles with rare cultural contexts.

But the evaluation is not reliable. The baselines are described only by name; nothing says StyleGAN2, DALL-E 2, or Original LDM were fine-tuned on the same enhanced DeepFashion+ data, while the proposed model gets two-phase training on LAION-5B plus fine-tuning. So the reported margins could just be fine-tuning. The ablation table makes it worse: 'Without Both Components' reports FID 11.30, while 'Original LDM' reports 10.55. If the ablation variant is the fine-tuned backbone, the baseline comparison is unfair; if it isn't, the numbers are internally inconsistent. Either way the reader cannot attribute the 7.80 FID to the proposed modules. The abstract says a 15% FID / 10% IS improvement, but Table I shows about 26% FID and 7.5% IS relative to Original LDM; headline and table don't match. On top of that, the semantic alignment metric in Table IV is the same cosine similarity used as a training objective in Eq. (7) and as a sample filter in Eq. (9), so the reported alignment gain is partially built into the method. There is no code, no dataset release, no error bars, and no training details.

Who is this for? Only a reader skimming fashion generation benchmarks might get a pointer, but there are better papers to read. This should be desk-rejected, not sent to review. If the authors later release code and data and fix the baseline and circularity problems, there might be a modest workshop paper in it, but in current form the evidence doesn't support the central claims.

Referee Report

5 major / 4 minor

Summary. The manuscript proposes a framework that combines an LLM-based prompt refinement module, a latent diffusion model (LDM), and a weak-supervision filtering module for cross-cultural fashion image generation. The method is trained in two phases: pre-training on LAION-5B and fine-tuning on an 'expanded DeepFashion+ dataset' enriched with culturally diverse styles. The authors report quantitative results (FID 7.80, IS 34.02) that they claim outperform StyleGAN2, DALL-E 2, and an unmodified LDM, along with human evaluation scores and a semantic alignment table.

Significance. If the central claims were fully supported, the paper would address a timely and useful problem: culturally inclusive fashion generation under weak supervision. The high-level idea of using LLMs to refine prompts and using a filtering module to handle noisy labels is plausible and worth investigating. However, the manuscript as written does not provide enough evidence for these claims: no code, model weights, dataset, training details, or error bars are released; the baseline comparisons are underspecified; and the reported semantic-alignment improvement is partly built into the training objective. The contribution is therefore currently at the level of a proposal rather than a validated method.

major comments (5)

[Abstract and Section IV-B, Table I] The abstract states that the method reduces FID by 15% and increases IS by 10%. From Table I, the proposed method's FID of 7.80 versus the Original LDM's 10.55 is a 26.1% reduction, and the IS increase from 31.64 to 34.02 is 7.5%. These headline numbers do not match the table, so the central quantitative claim is internally inconsistent.
[Section IV-A(c), Table I, Table II] The baseline setup is not specified. The paper does not state whether StyleGAN2, DALL-E 2, and Original LDM were fine-tuned on the same expanded DeepFashion+ data used for the proposed method. In addition, Table I reports Original LDM FID=10.55 while Table II reports 'Without Both Components' FID=11.30; if both are the same LDM backbone, the numbers should agree, and if they differ in fine-tuning protocol, the comparison and the ablation are not measuring the same thing. Without a matched fine-tuning protocol, the reported gains cannot be attributed to the proposed modules.
[Section III-C, Eqs. (7), (9), and Table IV] The prompt consistency loss in Eq. (7) trains the model using cosine similarity between CLIP text and image embeddings; Eq. (9) filters weakly labeled samples using the same cosine similarity; and Table IV reports 'semantic alignment' using a pre-trained CLIP model, i.e., the same measure. The reported alignment improvement is therefore partly enforced by construction. The authors should report alignment with a held-out metric that is not used in training or filtering.
[Section IV-A(a) and Section III-D] The 'expanded DeepFashion+ dataset' is never defined: its size, enrichment process, source images, annotation procedure, licensing, and train/test split are all absent. Every quantitative result depends on this dataset, so the experiments cannot be reproduced or verified. Similarly, the LLM, text encoder, image encoder, diffusion backbone, values of lambda_denoise, lambda_prompt, lambda_recon, threshold tau, number of diffusion steps, and random seeds are not reported.
[Tables I-III and Section IV-D] All quantitative results are single numbers with no error bars, confidence intervals, or multiple-seed runs. FID and IS are known to be sensitive to sample size and random seeds, so the claimed superiority over baselines is not statistically substantiated. The human evaluation also lacks statistical tests, inter-rater agreement, and details on participant selection and image sampling.

minor comments (4)

[Section IV-E(a)] The text refers to 'visual inspections' and 'qualitative assessments', but no figure or example generated image is included; the paper should provide sample grids illustrating cultural diversity and failure cases.
[Section II-A and References] Some citations do not support the claims they are attached to: for example, reference [8] is a hyperbolic geometric latent diffusion model for graph generation, but it is cited to support the claim that training datasets overrepresent specific styles and fail to capture global fashion diversity; a more relevant fashion-dataset citation is needed.
[Section III-C(a)] The module is called 'noise-aware filtering' but Eq. (9) scores samples by text-image cosine similarity and does not use noise information; the terminology should be aligned with the actual mechanism or the method should be extended accordingly.
[Section IV-A(b) and Table III] The human evaluation section should report whether the 50 participants were experts or laypersons, how the 30 images were selected per participant, and whether the score differences (e.g., 4.2 vs. 4.6 for relevance) are statistically significant.

Circularity Check

1 steps flagged · score 6.0 of 10

Semantic-alignment Table IV is the training objective; central FID/IS results are external and non-circular.

self definitional [Section III-C (Eqs. 7 and 9) and Section IV-E-c (Table IV)]
"Lprompt = E [1 − cos(Enctext(y), Encimg(ˆx))] , (7) ... s = cos(Enctext(y), Encimg(x)). (9) Only samples with s > τ, where τ is a threshold, are included in training. ... To quantify the alignment between textual prompts and generated images, we compute a semantic alignment score using a pre-trained CLIP model."

The paper's claimed evidence for prompt-image alignment (Table IV) is a cosine-similarity score between text and image embeddings. This is exactly the quantity that Eq. (7) trains the model to maximize (by minimizing 1−cos) and the quantity used in Eq. (9) to filter which samples enter training. No independent alignment measure is introduced in Table IV. Consequently, the reported improvement in semantic alignment is not an external test of the framework; it is the training objective and data-selection criterion re-evaluated on the model's outputs. The claim that the LLM-guided prompt refinement effectively bridges the gap therefore reduces by construction to the objective that was optimized, not to a free-standing prediction.

full rationale

The paper's headline FID/IS comparison is against external metrics and is not circular: FID and IS are computed from generated-image distributions independently of the training losses in Eqs. (6)-(10), and the method's lower FID/higher IS is a contingent empirical result. The same is true for the human evaluation. The one load-bearing circular step is Section IV-E-c/Table IV: the 'semantic alignment score using a pre-trained CLIP model' is the same cosine text-image similarity that Eq. (7) optimizes as a prompt-consistency loss and Eq. (9) uses to filter weakly labeled samples. Reporting this score as evidence that the LLM refinement 'bridges the gap between textual input and visual output' is therefore a self-referential evaluation; the metric is the objective. I found no self-citation chain: the closely related prior work [17] is by different authors, and no load-bearing argument rests on the present authors' own citations. Separately, but not circularity, the 'enhanced DeepFashion+ dataset' is never specified (size, enrichment procedure, licensing), and the baseline comparison does not state whether StyleGAN2/DALL-E 2/Original LDM were fine-tuned on the same data; the abstract's '15% FID / 10% IS' also does not match Table I (10.55→7.80 is ~26% FID; 31.64→34.02 is ~7.5% IS). These are reproducibility and fairness concerns, not reductions to inputs. Score 6 rather than higher because the main quantitative claim retains independent external support; only the alignment evidence reduces to the objective.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim depends on several unreported free parameters (threshold and loss weights), unverified domain assumptions about CLIP embeddings and dataset suitability, and two introduced artifacts (the expanded dataset and the filtering module) that are not released or independently validated.

free parameters (5)

threshold tau = not reported
Sample filtering threshold in Eq. (9), s > tau, controls how much weakly labeled data is used; value is not given.
lambda_denoise = not reported
Weight on denoising loss in Eq. (10).
lambda_prompt = not reported
Weight on prompt consistency loss in Eq. (10).
lambda_recon = not reported
Weight on reconstruction loss in Eq. (10).
LLM and LDM configuration = not reported
The pre-trained LLM and the latent diffusion model are not named, and their architectures, sizes, and decoding parameters are unspecified.

assumptions (4)

standard math Standard Gaussian diffusion forward and reverse process equations (1)-(6) hold as in prior latent diffusion literature.
The paper re-states standard diffusion equations without derivation or citation to the original DDPM/LDM work.
domain assumption CLIP embeddings (Enctext, Encimg) provide a valid semantic similarity measure for fashion images and cultural style.
Used in Eq. (7), Eq. (9), and Table IV; no validation that CLIP captures cultural diversity.
domain assumption LAION-5B and an 'expanded DeepFashion+' dataset are available and suitable for the pre-training and fine-tuning scheme.
The datasets are named but the expansion procedure and access details are not provided.
ad hoc to paper The denoising network p_theta is trained by minimizing L_total and converges as claimed.
No training curves, compute, or convergence details are given.

invented entities (2)

DeepFashion+ (expanded dataset enriched with culturally diverse styles)
purpose: Fine-tuning and evaluation dataset for cultural diversity
The dataset is not released, not specified in size or composition, and no URL is given.
Weak Supervision Filtering Module
purpose: Scores and filters weakly labeled samples by CLIP cosine similarity
It is a simple threshold operation; introduced as a new module but no theoretical or empirical justification beyond the reported tables.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Cross-Cultural Fashion Design via Interactive Large Language Models and Diffusion Models." pith.science (2026). https://pith.science/paper/Q5SZTIJC

@misc{pith2026250115571,
  author       = {Pith},
  title        = {Pith review of: Cross-Cultural Fashion Design via Interactive Large Language Models and Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q5SZTIJC}},
  note         = {Machine review of arXiv:2501.15571}
}

read the original abstract

Fashion content generation is an emerging area at the intersection of artificial intelligence and creative design, with applications ranging from virtual try-on to culturally diverse design prototyping. Existing methods often struggle with cultural bias, limited scalability, and alignment between textual prompts and generated visuals, particularly under weak supervision. In this work, we propose a novel framework that integrates Large Language Models (LLMs) with Latent Diffusion Models (LDMs) to address these challenges. Our method leverages LLMs for semantic refinement of textual prompts and introduces a weak supervision filtering module to effectively utilize noisy or weakly labeled data. By fine-tuning the LDM on an enhanced DeepFashion+ dataset enriched with global fashion styles, the proposed approach achieves state-of-the-art performance. Experimental results demonstrate that our method significantly outperforms baselines, achieving lower Frechet Inception Distance (FID) and higher Inception Scores (IS), while human evaluations confirm its ability to generate culturally diverse and semantically relevant fashion content. These results highlight the potential of LLM-guided diffusion models in driving scalable and inclusive AI-driven fashion innovation.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 13 canonical work pages

[17]

Interactive Fashion Content Generation Using LLMs and Latent Diffusion Models

K. S. I. Mantri and N. Sasikumar, “Interactive fashion c ontent generation using llms and latent diffusion models,” arXiv preprint arXiv:2306.05182, 2023. 5

work page Pith review arXiv 2023
[1]

On the robustness of latent diffusion models,

J. Zhang, Z. Xu, S. Cui, C. Meng, W. Wu, and M. R. Lyu, “On the robustness of latent diffusion models,” arXiv preprint arXiv:2306.08257, 2023

arXiv 2023
[2]

Smooth Diffusion: Crafting Smooth Latent Spaces in Diffusion Models

J. Guo, X. Xu, Y . Pu, Z. Ni, C. Wang, M. V asu, S. Song, G. Huan g, and H. Shi, “Smooth diffusion: Crafting smooth latent spaces in diffusion models,” arXiv preprint arXiv:2312.04410 , 2023

work page Pith review arXiv 2023
[3]

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
[5]

Planning with Logical Graph-based Language Model for Instruction Generation

F. Zhang, K. Jin, and H. H. Zhuo, “Planning with logical gr aph- based language model for instruction generation,” arXiv preprint arXiv:2308.13782, 2023

work page Pith review arXiv 2023
[6]

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

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

work page 2021
[7]

Eventber t: A pre- trained model for event correlation reasoning,

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

work page 2022
[8]

Hype rbolic geometric latent diffusion model for graph generation,

X. Fu, Y . Gao, Y . Wei, Q. Sun, H. Peng, J. Li, and X. Li, “Hype rbolic geometric latent diffusion model for graph generation,” arXiv preprint arXiv:2405.03188, 2024

arXiv 2024

Show all 31 references

[9]

Towards diffusion models for large-scale sea-ice modelling,

T. S. Finn, C. Durand, A. Farchi, M. Bocquet, and J. Brajar d, “Towards diffusion models for large-scale sea-ice modelling,” arXiv preprint arXiv:2406.18417, 2024

2024 arXiv
[10]

Rethinking visual d ependency in long-context reasoning for large vision-language models,

Y . Zhou, Z. Rao, J. Wan, and J. Shen, “Rethinking visual d ependency in long-context reasoning for large vision-language models, ” arXiv preprint arXiv:2410.19732, 2024

2024 arXiv
[12]

Fashion dna: Mer ging content and sales data for recommendation and article mapping,

C. Bracher, S. Heinz, and R. V ollgraf, “Fashion dna: Mer ging content and sales data for recommendation and article mapping,” arXiv preprint arXiv:1609.02489, 2016

2016 arXiv
[13]

Improving cross-modal alignment f or text- guided image inpainting,

Y . Zhou and G. Long, “Improving cross-modal alignment f or text- guided image inpainting,” in Proceedings of the 17th Conference of the European Chapter of the Association for Computational L inguistics, 2023, pp. 3445–3456

2023
[14]

Insectmamba: Inse ct pest classiﬁcation with state space model,

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

2024 arXiv
[15]

First: A million-entry dataset for text-driven fashion sy nthesis and design,

Z. Huang, Y . Li, D. Pei, J. Zhou, X. Ning, J. Han, X. Han, an d X. Chen, “First: A million-entry dataset for text-driven fashion sy nthesis and design,” arXiv preprint arXiv:2311.07414 , 2023

2023 arXiv
[16]

Unifashion: A uniﬁed vision-language model for multimodal fashion retrieval an d generation,

X. Zhao, Y . Zhang, W. Zhang, and X.-M. Wu, “Unifashion: A uniﬁed vision-language model for multimodal fashion retrieval an d generation,” arXiv preprint arXiv:2408.11305 , 2024

2024 arXiv
[18]

Snapmode: An intelligent and distributed large-scale fas hion image retrieval platform based on big data and deep generative adv ersarial network technologies,

N. Norouzi, R. Azmi, S. Saberi Tehrani Moghadam, and M. Z arvani, “Snapmode: An intelligent and distributed large-scale fas hion image retrieval platform based on big data and deep generative adv ersarial network technologies,” arXiv preprint arXiv:2204.03998 , 2022

2022 arXiv
[19]

Diffus ion models for generative outﬁt recommendation,

Y . Xu, W. Wang, F. Feng, Y . Ma, J. Zhang, and X. He, “Diffus ion models for generative outﬁt recommendation,” arXiv preprint arXiv:2402.17279, 2024

2024 arXiv
[20]

Woman-met al-white vs man-dress-shorts: Combining social, temporal and image signals to understand popularity of pinterest fashion boards,

S. K. Maity, A. Chaudhari, and A. Mukherjee, “Woman-met al-white vs man-dress-shorts: Combining social, temporal and image signals to understand popularity of pinterest fashion boards,” arXiv preprint arXiv:1812.07759, 2018

2018 arXiv
[21]

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,” arXiv preprint arXiv:2306.07377, 2023

2023 arXiv
[22]

A survey of large language models for european languages,

W. Ali and S. Pyysalo, “A survey of large language models for european languages,” arXiv preprint arXiv:2408.15040 , 2024

2024 arXiv
[23]

Weak to strong generaliz ation for large language models with multi-capabilities,

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

2024
[24]

How good are commercial large lang uage models on african languages?

J. Ojo and K. Ogueji, “How good are commercial large lang uage models on african languages?” arXiv preprint arXiv:2305.06530 , 2023

2023 arXiv
[25]

Cedille: A large autoregres sive french language model,

M. M¨ uller and F. Laurent, “Cedille: A large autoregres sive french language model,” arXiv preprint arXiv:2202.03371 , 2022

2022 arXiv
[26]

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,” arXiv preprint arXiv:2408.10441 , 2024

2024 arXiv
[27]

Towards robust ranker for text retrieval,

Y . Zhou, T. Shen, X. Geng, C. Tao, C. Xu, G. Long, B. Jiao, a nd D. Jiang, “Towards robust ranker for text retrieval,” in Findings of the Association for Computational Linguistics: ACL 2023 , 2023, pp. 5387– 5401

2023
[28]

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
[29]

Large lan guage models in bioinformatics: applications and perspectives,

J. Liu, M. Y ang, Y . Y u, H. Xu, K. Li, and X. Zhou, “Large lan guage models in bioinformatics: applications and perspectives, ” arXiv preprint arXiv:2401.04155, 2024

2024 arXiv
[30]

Llama-reg: Using llama 2 for unsuperv ised medical image registration,

M. Ma and Y . Y ang, “Llama-reg: Using llama 2 for unsuperv ised medical image registration,” arXiv preprint arXiv:2405.18774 , 2024

2024 arXiv
[31]

A precis of language models are not models of l anguage,

C. V eres, “A precis of language models are not models of l anguage,” arXiv preprint arXiv:2205.07634 , 2022

2022 arXiv
[32]

Modelling language,

J. Grindrod, “Modelling language,” arXiv preprint arXiv:2404.09579 , 2024

2024
[33]

Beyond the limitations of any imaginable mechanism: large language models and psyc holin- guistics,

C. Houghton, N. Kazanina, and P . Sukumaran, “Beyond the limitations of any imaginable mechanism: large language models and psyc holin- guistics,” arXiv preprint arXiv:2303.00077 , 2023. 6

2023 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [17]

Interactive Fashion Content Generation Using LLMs and Latent Diffusion Models

K. S. I. Mantri and N. Sasikumar, “Interactive fashion c ontent generation using llms and latent diffusion models,” arXiv preprint arXiv:2306.05182, 2023. 5

work page Pith review arXiv 2023

[2] [1]

On the robustness of latent diffusion models,

J. Zhang, Z. Xu, S. Cui, C. Meng, W. Wu, and M. R. Lyu, “On the robustness of latent diffusion models,” arXiv preprint arXiv:2306.08257, 2023

arXiv 2023

[3] [2]

Smooth Diffusion: Crafting Smooth Latent Spaces in Diffusion Models

J. Guo, X. Xu, Y . Pu, Z. Ni, C. Wang, M. V asu, S. Song, G. Huan g, and H. Shi, “Smooth diffusion: Crafting smooth latent spaces in diffusion models,” arXiv preprint arXiv:2312.04410 , 2023

work page Pith review arXiv 2023

[4] [3]

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

[5] [5]

Planning with Logical Graph-based Language Model for Instruction Generation

F. Zhang, K. Jin, and H. H. Zhuo, “Planning with logical gr aph- based language model for instruction generation,” arXiv preprint arXiv:2308.13782, 2023

work page Pith review arXiv 2023

[6] [6]

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

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

work page 2021

[7] [7]

Eventber t: A pre- trained model for event correlation reasoning,

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

work page 2022

[8] [8]

Hype rbolic geometric latent diffusion model for graph generation,

X. Fu, Y . Gao, Y . Wei, Q. Sun, H. Peng, J. Li, and X. Li, “Hype rbolic geometric latent diffusion model for graph generation,” arXiv preprint arXiv:2405.03188, 2024

arXiv 2024