Towards Vision-Language-Garment Models for Web Knowledge Garment Understanding and Generation

Gordon Wetzstein; Guandao Yang; Jan Ackermann; Kiyohiro Nakayama; Tong Wu

REVIEW 3 major objections 5 minor 18 references

Towards Vision-Language-Garment Models for Web Knowledge Garment Understanding and Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A vision-language model fine-tuned on garment tokens transfers web-scale reasoning to unseen styles and prompts, beating task-specific baselines on synthetic out-of-distribution tests.

desk verdict Plausible incremental extension of AIpparel with a useful prompt dataset, but the OOD transfer claim is undercut by evaluation sets generated from the same pipeline as training. read the letter →

arxiv 2506.05210 v2 pith:HBQEB4PC submitted 2025-06-05 cs.CV

Jan Ackermann , Kiyohiro Nakayama , Guandao Yang , Tong Wu , Gordon Wetzstein This is my paper

classification cs.CV

keywords vision-languagemodelsgarmentgenerationsewingpatternszero-shotgeneralizationout-of-distributiontransfermultimodalreasoningfine-tuning

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

The paper sets out to show that a vision-language model pre-trained on web-scale data can carry its general reasoning into the narrow task of garment generation. It introduces VLG, which reads a garment image or a natural-language prompt and emits a sewing-pattern token sequence, and it tests whether this fine-tuned model generalizes to styles, backgrounds, scales, and phrasings it did not see in training. On its synthetic benchmark, VLG reports far higher success rates than the task-specific baselines, which the paper takes as evidence that pre-training, not task-specific architecture, provides out-of-distribution robustness. The paper also finds that prompt diversity matters more than texture diversity and that accuracy keeps improving with dataset size, while noting that underspecified prompts and very dissimilar real images remain hard.

What carries the argument

The load-bearing mechanism is a fine-tuned vision-language model that autoregressively predicts garment sewing-pattern tokens from either an image or a text prompt. The vision encoder and cross-attention layers stay frozen so the model retains its web-scale visual and linguistic knowledge, and only the prediction head is tuned to emit tokens from a garment-specific codebook. The argument is carried by a synthetic data pipeline: parametric sewing-pattern garments are rendered, converted to edge maps, textured by a conditioned diffusion model, and paired with naturalistic vague prompts generated by a large language model, producing variation in background, position, scale, appearance, and phrasing. That pipeline supplies both the training set and the synthetic tests, and it is what lets the paper measure reasoning transfer dimension by dimension.

What would settle it

Take 1,000 real garment photographs (not renders from the synthetic pipeline), obtain ground-truth sewing patterns or expert sewability ratings, and compare VLG's garment accuracy with a fine-tuned garment-specific baseline. If the margin on these real photos is much smaller than the margins reported on synthetic tests, the transfer claim is an artifact of train/test distribution overlap rather than web-scale reasoning transfer.

Watch

Extended reading notes

Core claim

In the paper's telling, the discovery is that web-scale pre-training is what makes a garment-generation model robust to out-of-distribution inputs. VLG keeps the vision encoder and cross-attention of a pre-trained vision-language model frozen, fine-tunes a prediction head to output discrete sewing-pattern tokens, and then beats garment-specific baselines on synthetic tests that change backgrounds, positions, scales, textures, and prompt wording, landing at roughly 74 to 88 percent success where the baselines land much lower. The same model degrades on real images as those images move away from the training distribution, and it improves on vague prompts when familiar garment words appear in new contexts. Dataset ablations show that adding natural-language prompt variety boosts text accuracy more than adding textured renders does, and that accuracy keeps rising with training-set size. The paper presents these results as evidence that general vision-language reasoning transfers into specialized fashion-design tasks.

Load-bearing premise

The out-of-distribution tests are synthetic and generated by the same pipeline that made the training data, so the measured margins may reflect overlap with training rather than true transfer to real garments and naturally written prompts.

Editorial extensions

If this is right

If pre-training is the source of robustness, the same fine-tuning recipe should work for other structured-output design domains whose artifacts can be tokenized, not just garments.
Because accuracy keeps rising with dataset size, collecting more garment-prompt pairs is the most direct route to a stronger model.
Prompt diversity, not texture diversity, is what improves language grounding, so future data collection should prioritize varied natural phrasings over varied appearances.
The model's real-photo performance is bounded by similarity to its training distribution, so deployment needs either broader data coverage or explicit rejection of far out-of-distribution inputs.

Reading between the lines

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

Because training and test sets come from the same rendering, language-model, and texture pipeline, the reported margins likely overstate transfer to real consumer photos and naturally written prompts; a real-photo benchmark is the missing experiment.
The frozen vision encoder means visual generalization is inherited from the backbone, not learned from garment data; unfreezing it and measuring the change would isolate where the transfer actually comes from.
The benchmark perturbs one factor at a time, but real garment photos shift background, scale, position, and appearance simultaneously, so combined perturbations would probably compress the reported margins.
The model's sensitivity to familiar words suggests an interactive design system that starts from vague language and asks clarifying questions, since accuracy improves as prompts accumulate specific garment vocabulary.

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 a modest, honest extension of AIpparel. The new pieces are a GPT-4o-generated prompt corpus and a synthetic textured garment dataset, plus an OOD evaluation study. The central claim—that pre-training transfers web-scale reasoning to garment pattern generation—is plausible but not demonstrated, because the evaluation data is generated by the same pipeline used for training.

What the paper does well: the prompt dataset is a genuinely useful resource. The ablation in Table 1 is informative: adding varied text improves text accuracy with only a modest dip in image-based metrics. The convergence and data-scaling figures are standard but cleanly done. The paper is also transparent about limitations; Sec. 5 explicitly says generalization to OOD scenarios is still a challenge, and Fig. 3 honestly shows degradation as inputs move away from the training distribution.

The soft spots are real and central. The OOD test sets are constructed by applying variations to GarmentCode renders, using ControlNet with GPT-4o prompts for textures, and using GPT-4o for the held-out text set. That means the 'unseen' inputs sit in the same distribution family as the training data. Manual curation of the evaluation set (Sec. 3.1) with no documented protocol adds selection-bias risk. Fig. 2 has no error bars or statistical tests, and exact test set sizes are missing. The paper's own Fig. 3 undercuts the OOD transfer story: performance drops steadily as inputs become more distant, which is exactly what you'd expect from a model fitting the training distribution.

I want to stress that the paper doesn't oversell. It says 'preliminary results' and 'highlighting the potential.' The headline claim is under-supported, but the underlying engineering is reasonable and the dataset contribution stands on its own.

Who this is for: researchers in garment generation or multimodal transfer who want a quick, honest look at whether fine-tuning a VLM on sewing patterns is promising, and who might use the new prompt dataset. It's not a definitive study. I'd send it to peer review because the community benefits from this kind of incremental data contribution and transparent exploration, but I'd make acceptance conditional on releasing code and data, adding error bars, and evaluating on a genuinely external OOD benchmark. As it stands, conditionally acceptable.

Recommendation: engage with it, but ask for the external evaluation before you trust the headline.

Referee Report

3 major / 5 minor

Summary. VLG is a vision-language-garment model built by fine-tuning a pre-trained vision-language backbone with AIpparel's garment tokenization to predict sewing patterns from text and images. The authors introduce a GPT-4o-generated prompt dataset and a ControlNet-based textured garment image dataset, then evaluate the model on variations in background, position, scale, appearance, and language, comparing with DressCode, SewFormer, and SewFormer-FT. The central claim is that robust pre-training enables VLG to transfer web-scale reasoning to unseen garment styles and prompts, with quantitative success rates reported in Fig. 2 and supporting ablations in Table 1 and Figs. 5-6.

Significance. If the transfer claim were established, this would be a useful empirical demonstration that vision-language foundation models can generalize to garment pattern generation, and the released datasets would be a resource for the community. The paper is honestly worded as a preliminary study and explicitly acknowledges performance degradation away from the training distribution, which is a strength. However, the paper does not ship code or model weights, and the evaluation design currently does not rule out distribution overlap as the explanation for the reported OOD success; the contribution is therefore not yet at the level of a definitive demonstration of web-scale knowledge transfer.

major comments (3)

[Sec. 3.1 and Sec. 4 (Fig. 2)] The out-of-distribution evaluation sets are generated by the same GarmentCode/GPT-4o/ControlNet pipeline used for training: Sec. 3.1 describes GPT-4o prompt generation, Canny/ControlNet texture synthesis, and manual curation, and Sec. 4 then evaluates 'visual generalization' via background, translation, zoom, and random-texture variations plus a GPT-4o-generated held-out text set. Because the test distribution overlaps the training-generation pipeline, the high relative success rates in Fig. 2 can be explained by distribution overlap rather than transfer of web-scale reasoning. To support the central claim, the authors should evaluate on data produced by a genuinely independent process, such as real garment photographs and human-written prompts, or provide a documented protocol that guarantees the test stimuli are outside the training distribution (e.g., unseen garment classes, a different texture synthesizer, or a separate prompt-generation model).
[Sec. 4, Metrics and Fig. 2] The quantitative comparison reports single-point success rates without error bars, confidence intervals, significance tests, or exact evaluation-set sizes. Several differences between VLG and the baselines are only a few percentage points (for example, 74.4% vs. 78.6% in Scale Generalization), so without variance estimates it is unclear whether the reported advantage is statistically meaningful. The authors should report the number of evaluation examples per reasoning dimension, the number of seeds or bootstrap resamples, and appropriate uncertainty measures.
[Sec. 3.1, Sec. 4 (Limits of Reasoning)] The 'held-out text test set' is produced with GPT-4o, the same model used to generate the training prompts, and the visual OOD variants derive from the same GarmentCode and ControlNet pipeline. This makes the language-grounding scores in Fig. 2 a measure of reproducing GPT-4o's phrasing distribution as much as a measure of natural-language understanding. In addition, the statement that 'models without pre-training struggle with OOD data, whereas our VLG benefits significantly from robust pre-training' is partly undercut by the authors' own Fig. 3, which shows degradation with 'distance to training distribution' without defining that distance. A quantified distance metric and a test set spanning a wide, well-characterized distance range are needed before the main claim can be accepted.

minor comments (5)

[Sec. 4, Metrics] There is a typo in 'quantitavely evaluate' (should be 'quantitatively evaluate').
[Fig. 2 caption] The caption lists the sub-dimensions but does not define what 'success rate' means for each task (e.g., threshold on garment accuracy, vertex L2, or human evaluation); please provide precise definitions and the number of samples per bar.
[Sec. 3.1] The text says backgrounds are masked white for consistency, but Sec. 4 evaluates 'background variations' as an OOD dimension; clarify how the background-variation test set differs from the masked-white training images.
[Sec. 3.2] The description of architecture and training states that the vision encoder and cross-attention modules are frozen, but it does not specify exactly which modules are trainable beyond the regression head; please provide a table of trainable vs. frozen parameters to enable reproduction.
[References] Reference [6] is cited as 'GarmentCodeData' in Sec. 4 but listed as 'Garmentcodedata' in the reference list; normalize the capitalization and ensure all references are cited in text.

Circularity Check

2 steps flagged · score 4.0 of 10

OOD generalization claim rests on evaluation sets produced by the same GarmentCode/GPT-4o/ControlNet pipeline used for training, so reported transfer may reflect in-distribution fit rather than measured web-scale knowledge transfer.

fitted input called prediction [Sec. 3.1 (Garment Prompt Dataset) and Sec. 4 (Garment Reasoning Analogous)]
"To bridge this gap, we leverage GPT-4o's natural language generation abilities to create a dataset with diverse, context-rich, and realistically ambiguous prompts ... For language grounding, we utilize a held-out text test set."

The only language-grounding evaluation is a held-out split of the same GPT-4o prompt-generation procedure used to build the training 'Garment Prompt Dataset.' A model fine-tuned on GPT-4o-generated prompts is then scored on GPT-4o-generated held-out prompts, so the reported language-grounding success measures fit to the GPT-4o prompting style rather than a transfer of web-scale reasoning to genuinely natural human prompts. The paper provides no evidence that the held-out text set was independently authored or drawn from real human interactions.
fitted input called prediction [Sec. 3.1 (Synthetic Textured Garment Dataset) and Sec. 4 (Garment Reasoning Analogous)]
"Our method begins with rendered garment images and utilizes the Canny Edge Detector to derive control masks. GPT-4o [11] is then queried to generate plausible textual descriptions for textures, which are subsequently synthesized using ControlNet [17] into neural-rendered images. ... For evaluation, we refine some synthesized images using a Virtual Try-On tool [2]. Given the limitations of such tools in geometry preservation, we manually curate the final dataset."

The 'out-of-distribution' visual evaluation in Sec. 4 is a synthetic dataset of backgrounds, translations, zooms, and random textures applied to the same GarmentCode renders and ControlNet textures from Sec. 3.1, and Sec. 4 says 'We evaluate predictions on images discussed in Sec. 3.1.' The evaluation therefore perturbs the training-generation distribution rather than sampling genuinely external garment data. The paper's own Fig. 3 caption states 'Pattern predictions degrade as input becomes increasingly distant from the training distribution,' confirming the model tracks the training pipeline; the reported visual/position/scale/appearance generalization can be explained by distribution overlap.

full rationale

This paper has no mathematical derivation chain, so circularity, if present, is in the construction of the evaluation data. The training data (GarmentCode renders, GPT-4o-generated prompts, ControlNet textures) and the held-out/OOD test data are generated by the same pipeline: Sec. 3.1 states that the two contributed datasets are 'designed explicitly to train and benchmark' the reasoning abilities, and Sec. 4 evaluates on 'images discussed in Sec. 3.1' plus a 'held-out text test set' from the GPT-4o prompt-generation process. Consequently, the central claim of web-scale reasoning transfer to unseen garments and prompts is measured on distributions that overlap with the fine-tuning distribution. This is a real, but partial, circularity: the model architecture, tokenization, baselines, and ablations are independent contributions, and the paper honestly notes in Sec. 5 that 'limited inductive biases imply challenges in generalizing to out-of-distribution scenarios.' No load-bearing self-citation was found; the AIpparel tokenization (Ref. [10]) is a representation choice rather than an unverified premise. Score 4 reflects partial data-construction circularity confined to the OOD evaluation, not a fully circular derivation.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the choice of AIpparel tokenization, on synthetic data generated by GPT-4o and ControlNet, and on manual curation of evaluation images. The only numeric inputs are hand-chosen training hyperparameters. No new physical entities are postulated.

free parameters (2)

Peak learning rate = 6e-5
Hand-chosen in Sec. 3.2 for fine-tuning; not fitted to data, but it controls convergence and the reported results.
Training epochs = 20
Hand-chosen in Sec. 3.2; all ablation models use identical compute budgets, so epoch count interacts with the number of data types.

assumptions (3)

domain assumption AIpparel garment tokenization is a sufficient representation of garments for measuring generation quality.
Sec. 3.2 states 'we utilize AIpparel's garment tokenization scheme to represent garments'; if the tokenization is lossy or biased, all accuracy metrics inherit that bias.
ad hoc to paper GPT-4o-generated prompts and ControlNet textures approximate real human garment descriptions and appearances.
Sec. 3.1 builds the prompt and texture datasets with GPT-4o and ControlNet; the benchmark distribution is defined by these synthetic inputs, not by independently collected human data.
ad hoc to paper Manual curation of evaluation images preserves an unbiased sample of the target distribution.
Sec. 3.1 says 'we manually curate the final dataset, ensuring suitability for rigorous multimodal reasoning evaluation'; curation can select favorable examples and inflate generalization scores.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Vision-Language-Garment Models for Web Knowledge Garment Understanding and Generation." pith.science (2026). https://pith.science/paper/HBQEB4PC

@misc{pith2026250605210,
  author       = {Pith},
  title        = {Pith review of: Towards Vision-Language-Garment Models for Web Knowledge Garment Understanding and Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HBQEB4PC}},
  note         = {Machine review of arXiv:2506.05210}
}

read the original abstract

Multimodal foundation models have demonstrated strong generalization, yet their ability to transfer knowledge to specialized domains such as garment generation remains underexplored. We introduce VLG, a vision-language-garment model that synthesizes garments from textual descriptions and visual imagery. Our experiments assess VLG's zero-shot generalization, investigating its ability to transfer web-scale reasoning to unseen garment styles and prompts. Preliminary results indicate promising transfer capabilities, highlighting the potential for multimodal foundation models to adapt effectively to specialized domains like fashion design.

Figures

Figures reproduced from arXiv: 2506.05210 by the authors.

**Figure 1.** Vision-Language-Garment Models. We tune existing vision-language models for garment-specific tasks to predict sewing patterns from multimodal (textual and visual) inputs. 1. The introduction of VLG, a multimodal garment generation model designed to leverage generalized visionlanguage reasoning for garment-specific tasks. 2. An evaluation of the reasoning capabilities of VLG against state-of-the-art methods, using… view at source ↗

**Figure 2.** Reasoning Transfer Evaluation. Performance comparison of our proposed model (red) against DressCode [3] (blue), SewFormer [8] (orange) and Sewformer-FT (green), across various reasoning dimensions. ing [13, 16] and domain-specific fine-tuning [7, 15, 18], though fine-tuning often risks catastrophic forgetting. Notably, robotics-focused approaches such as RT-2 [14] and OpenVLA [12] successfully retain general reaso… view at source ↗

**Figure 4.** Textual Generalization. Increasing specificity in textual prompts leads to more accurate garment pattern predictions. in [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: Validation Loss by Iteration. Garment accuracy steadily improves, and Vertex L2 loss decreases over training iterations, indicating successful model convergence. Ablations Given that we fine-tune a vision-language model on a novel domain with newly introduced tokens, …

**Figure 6.** Figure 6: Validation Loss by Data Size. Increasing the training dataset size leads to improved garment accuracy and reduced vertex prediction error, highlighting the importance of dataset scale. generalization. All models were trained with identical compute budgets; thus, intr…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 17 canonical work pages

[1]

Chat- garment: Garment estimation, generation and editing via large language models

Siyuan Bian, Chenghao Xu, Yuliang Xiu, Artur Grigorev, Zhen Liu, Cewu Lu, Michael J Black, and Yao Feng. Chat- garment: Garment estimation, generation and editing via large language models. CVPR, 2025. 1

work page 2025
[2]

Improving diffusion models for au- thentic virtual try-on in the wild

Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for au- thentic virtual try-on in the wild. ECCV, 2024. 2

work page 2024
[3]

Dresscode: Autoregressively sewing and gen- erating garments from text guidance

Kai He, Kaixin Yao, Qixuan Zhang, Jingyi Yu, Lingjie Liu, and Lan Xu. Dresscode: Autoregressively sewing and gen- erating garments from text guidance. ACM Transactions on Graphics, 2024. 1, 2

work page 2024
[4]

Neuraltailor: Recon- structing sewing pattern structures from 3d point clouds of garments

Maria Korosteleva and Sung-Hee Lee. Neuraltailor: Recon- structing sewing pattern structures from 3d point clouds of garments. ACM Transactions on Graphics, 2022. 1

work page 2022
[5]

Garment- code: Programming parametric sewing patterns.ACM Trans- actions on Graphics, 2023

Maria Korosteleva and Olga Sorkine-Hornung. Garment- code: Programming parametric sewing patterns.ACM Trans- actions on Graphics, 2023. 1

work page 2023
[6]

Garmentcodedata: A dataset of 3d made-to-measure garments with sewing pat- terns

Maria Korosteleva, Timur Levent Kesdogan, Fabian Kem- per, Stephan Wenninger, Jasmin Koller, Yuhan Zhang, Mario Botsch, and Olga Sorkine-Hornung. Garmentcodedata: A dataset of 3d made-to-measure garments with sewing pat- terns. ECCV, 2024. 3

work page 2024
[7]

Visual instruction tuning

Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 2

work page 2023
[8]

Towards garment sewing pattern reconstruc- tion from a single image

Lijuan Liu, Xiangyu Xu, Zhijie Lin, Jiabin Liang, and Shuicheng Yan. Towards garment sewing pattern reconstruc- tion from a single image. ACM Transactions on Graphics ,

work page

Show all 18 references

[9]

Llama 2: Open foundation and fine-tuned chat models, 2023

LLaMA-Team. Llama 2: Open foundation and fine-tuned chat models, 2023. 1

2023
[10]

Aipparel: A large multimodal generative model for digital garments

Kiyohiro Nakayama, Jan Ackermann, Timur Levent Kes- dogan, Yang Zheng, Maria Korosteleva, Olga Sorkine- Hornung, Leonidas Guibas, Guandao Yang, and Gordon Wetzstein. Aipparel: A large multimodal generative model for digital garments. CVPR, 2025. 1, 2, 3

2025
[11]

Gpt-4 technical report, 2024

OpenAI. Gpt-4 technical report, 2024. 1, 2

2024
[12]

Openvla: An open-source vision- language-action model

OpenVLA-Team. Openvla: An open-source vision- language-action model. 2024. 2, 3

2024
[13]

Hugginggpt: solving ai tasks with chatgpt and its friends in hugging face

Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: solving ai tasks with chatgpt and its friends in hugging face. NeurIPS,
[14]

Rt-2: Vision-language-action models transfer web knowledge to robotic control

RT-2 Team. Rt-2: Vision-language-action models transfer web knowledge to robotic control. 2023. 2, 3

2023
[15]

Cambrian-1: A fully open, vision-centric exploration of multimodal llms,

Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms,
[16]

Visual chatgpt: Talking, drawing and editing with visual foundation models, 2023

Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models, 2023. 2

2023
[17]

Adding conditional control to text-to-image diffusion models

Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2

2023
[18]

Minigpt-4: Enhancing vision-language understanding with advanced large language models

Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mo- hamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. 2023. 2

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Chat- garment: Garment estimation, generation and editing via large language models

Siyuan Bian, Chenghao Xu, Yuliang Xiu, Artur Grigorev, Zhen Liu, Cewu Lu, Michael J Black, and Yao Feng. Chat- garment: Garment estimation, generation and editing via large language models. CVPR, 2025. 1

work page 2025

[2] [2]

Improving diffusion models for au- thentic virtual try-on in the wild

Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for au- thentic virtual try-on in the wild. ECCV, 2024. 2

work page 2024

[3] [3]

Dresscode: Autoregressively sewing and gen- erating garments from text guidance

Kai He, Kaixin Yao, Qixuan Zhang, Jingyi Yu, Lingjie Liu, and Lan Xu. Dresscode: Autoregressively sewing and gen- erating garments from text guidance. ACM Transactions on Graphics, 2024. 1, 2

work page 2024

[4] [4]

Neuraltailor: Recon- structing sewing pattern structures from 3d point clouds of garments

Maria Korosteleva and Sung-Hee Lee. Neuraltailor: Recon- structing sewing pattern structures from 3d point clouds of garments. ACM Transactions on Graphics, 2022. 1

work page 2022

[5] [5]

Garment- code: Programming parametric sewing patterns.ACM Trans- actions on Graphics, 2023

Maria Korosteleva and Olga Sorkine-Hornung. Garment- code: Programming parametric sewing patterns.ACM Trans- actions on Graphics, 2023. 1

work page 2023

[6] [6]

Garmentcodedata: A dataset of 3d made-to-measure garments with sewing pat- terns

Maria Korosteleva, Timur Levent Kesdogan, Fabian Kem- per, Stephan Wenninger, Jasmin Koller, Yuhan Zhang, Mario Botsch, and Olga Sorkine-Hornung. Garmentcodedata: A dataset of 3d made-to-measure garments with sewing pat- terns. ECCV, 2024. 3

work page 2024

[7] [7]

Visual instruction tuning

Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 2

work page 2023

[8] [8]

Towards garment sewing pattern reconstruc- tion from a single image

Lijuan Liu, Xiangyu Xu, Zhijie Lin, Jiabin Liang, and Shuicheng Yan. Towards garment sewing pattern reconstruc- tion from a single image. ACM Transactions on Graphics ,

work page