Pith. sign in

REVIEW 3 major objections 5 minor 42 references

HiGarment: Cross-modal Harmony Based Diffusion Model for Flat Sketch to Realistic Garment Image

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 The paper claims that HiGarment, a diffusion model with fabric retrieval and harmonized cross-attention, generates realistic garment images from flat sketches and text, and outperforms five baselines on all reported metrics.

desk verdict A solid empirical package for a new garment-generation task, but the HCA equations don't yet support the claimed modality-balancing behavior. read the letter →

arxiv 2505.23186 v2 pith:KIBIUS6G submitted 2025-05-29 cs.CV

Junyi Guo , Jingxuan Zhang , Fangyu Wu , Huanda Lu , Qiufeng Wang , Wenmian Yang , Eng Gee Lim , Dongming Lu This is my paper
classification cs.CV
keywords flatsketchtorealisticgarmentimagesynthesisdiffusionmodelcross-modalharmonizedattentionfabricretrievalmulti-modaldatasettext-guidedgenerationfashionproduction
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 introduces a new task, Flat Sketch to Realistic Garment Image (FS2RG), where the input is a flat technical sketch plus a text prompt and the output is a realistic photo of the finished garment. It argues that existing diffusion adapters fail on this task because they align image and text features only when the two modalities agree, so fabric attributes carried only by text get ignored and conflicts cannot be resolved. The proposed HiGarment adds a fabric-retrieval module that injects fabric samples into both visual and textual embeddings, plus a harmonized cross-attention mechanism that weights sketch and text information with a scalar derived from their cosine similarity. On the new MMDGarment dataset, the paper reports gains over the strongest baseline of 3.7% in CLIPScore, 20.3% in FID, and 14.5% in LPIPS, along with the highest expert and non-expert human ratings.

What carries the argument

The load-bearing object is the harmonized cross-attention mechanism (HCA) operating on features produced by the multi-modal semantic enhancement (MMSE) module. MMSE uses a Q-Former over a fabric label and a retrieved fabric sample, then cross-attends the resulting fabric feature into the CLIP image embedding and text embedding to form enhanced features $v'$ and $t'$. HCA builds the query $Q$ from $v'$, concatenates $v'$ and $t'$ into the keys and values, and scales the standard attention output by the weight $\alpha = \lambda + (1-\lambda)\sigma(s)$, where $\sigma(s)$ is the sigmoid of the cosine similarity between the original image and text embeddings and $\lambda$ is set to 0.6. This weight is meant to be the single control dial that shifts generation between sketch-aligned and text-biased outputs at each denoising step.

What would settle it

Run the released model on a fixed sketch paired with a conflicting text prompt (for example, a red sketch with the prompt 'blue') and sweep alpha from 1 down to 0.6. If the output color does not move monotonically toward the text, or if removing the base Stable Diffusion text cross-attention leaves the output unchanged at low alpha, then the claimed text-bias control does not come from the harmonized cross-attention mechanism.

Watch

Extended reading notes

Core claim

The central discovery claimed is that a garment-specific cross-modal architecture can make a pretrained Stable Diffusion model respect both precise sketch structure and text-only fabric attributes. By retrieving a fabric sample from a vector database, processing it with a Q-Former, and injecting the resulting fabric feature into both the image and text embeddings, HiGarment gains fine-grained material control that standard CLIP alignment lacks. A harmonized cross-attention layer then computes the cosine similarity between the enhanced image and text embeddings, maps it to a weight alpha between 0.6 and 1, and scales the concatenated attention output, so high similarity favors sketch-aligned rendering and low similarity favors text-guided modification. The paper reports state-of-the-art results on the FS2RG task and attributes the gains to these two modules through ablations.

Load-bearing premise

The paper assumes that a single number computed from sketch-text similarity really chooses between sketch and text guidance during generation, but the equation shown multiplies both branches by the same number, so the effect must rely on an extra pathway in the frozen base model that is not described.

Editorial extensions

If this is right

  • Flat sketches can serve as production-stage input for garment prototyping, with localized edits such as pockets, hoods, and fabric changes specified by text instead of redrawing the sketch.
  • A relatively small adapter trained on a garment-specific dataset can outperform general-purpose image-text adapters on this specialized synthesis task, so domain-specific representation modules matter more than generic alignment.
  • The released MMDGarment dataset with flat sketches, close-ups, a fabric database, and text annotations enables reproducible benchmarking of FS2RG, a task that previously had no public benchmark.
  • Because alpha can be clamped manually, users can choose image-biased output (alpha near 1) or text-biased output (alpha near 0.6), providing a single dial for fidelity-versus-modification trade-offs.
  • The authors acknowledge that tiny logos and intricate details are often lost during denoising, so FS2RG is not yet a complete replacement for physical sampling.

Reading between the lines

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

  • If the HCA equation is taken literally, alpha scales both the sketch and text attention outputs together, so the claimed modality selection may actually be carried by an unspecified residual pathway in the frozen Stable Diffusion cross-attention; a direct ablation of that pathway would settle the question.
  • The same two-module recipe could transfer to other sketch-plus-text production tasks, such as architectural drawings or technical product renders, where a reference image supplies structure and text supplies material or finish attributes.
  • Because the paper reports only aggregate metrics on the test set, a per-attribute breakdown (fabric versus color versus structure) would show which modality the harmony mechanism actually favors when they conflict.
  • The fabric-retrieval idea generalizes: any attribute that is hard to name but easy to show, such as leather grain, knit density, or sheen, could be retrieved from a small visual dictionary instead of relying on text-only CLIP features.
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

3 major / 5 minor

Summary. The paper introduces a new task, Flat Sketch to Realistic Garment Image (FS2RG), in which a diffusion model generates a realistic garment image from a flat sketch and a text prompt. To support the task, the authors collect and release MMDGarment, a dataset of 20,151 garment images with flat sketches, close-ups, and detailed text annotations, plus a 150-entry fabric database. The proposed method HiGarment combines a Multi-Modal Semantic Enhancement (MMSE) module, which augments visual and textual features using retrieved fabric samples, and a Harmonized Cross-Attention (HCA) mechanism, which scales the cross-attention output by a similarity-dependent factor alpha to allegedly balance image and text guidance. Experiments compare HiGarment with IP-Adapter, Uni-ControlNet, Versatile Diffusion, DEADiff, and SSR-Encoder on the MMDGarment test set, reporting improved CLIPScore, FID, and LPIPS, as well as higher human and MLLM ratings. The paper also includes ablations that remove MMSE, HCA, and the MMDGarment training data.

Significance. If the claimed results hold, the FS2RG task and the MMDGarment dataset are useful contributions to garment-oriented image generation, and HiGarment would represent a new state of the art on that task. The strengths of the paper include the release of code and dataset, the breadth of baseline comparisons, and the ablation study that separates module contributions from dataset-scale effects. The main concern is that the central mechanism claimed for HCA, dynamic balancing between image and text modalities, is not actually represented by the equations as written; the paper leaves unspecified how the scaled attention output is combined with the base Stable Diffusion text-conditioning path. This needs to be addressed before the methodological claims can be accepted. The empirical evaluation also lacks statistical reliability, which weakens the claim of superiority over the best baseline.

major comments (3)
  1. [Sec. 4.2.3, Eq. (10)] The HCA mechanism as written cannot implement the claimed image/text balancing. Since K and V are concatenations of v' and t' before the softmax, z = alpha * Softmax(QK^T/sqrt(d)) V is a single gated attention over a mixed multimodal context; scaling by alpha attenuates or amplifies the combined output rather than choosing between image and text. The claimed behavior only follows if z is injected alongside an independent text-only pathway whose output is unaffected by alpha (for example, if Eq. (10) is added to the base SD cross-attention output rather than replacing it), but the paper states only that z is 'integrated into the pre-trained Stable Diffusion UNet by injecting it into each cross-attention layer.' This is load-bearing because the HCA ablation and the method's novelty rest on the balancing property. The authors should specify the exact combination rule, provide the corresponding equations, and report an ablation where the base text pathway is removed or kept constant while alpha varies.
  2. [Sec. 5.3, Tables 1-3] The state-of-the-art claim is not supported by statistical evidence. Tables 1 and 3 report single-run values without error bars, repeated seeds, or significance tests, so the reported margins over IP-Adapter (3.7% in CLIPScore, 20.3% in FID, 14.5% in LPIPS) may be within run-to-run variance. The user study in Table 2 and Supp. Sec. 9 uses only 28 pairs for the MLLM evaluation and aggregate scores from 33 participants without inter-rater agreement or a significance test. I recommend reporting mean and standard deviation over at least three seeds for the main quantitative comparisons, adding a permutation or paired test for the baseline differences, and reporting per-pair variance or a Wilcoxon signed-rank test for the human and MLLM evaluations.
  3. [Sec. 5.5, Eq. (6) and Fig. 7] The choice of lambda = 0.6 is made by visual inspection of Figure 7, yet alpha is the mechanism through which HCA is claimed to balance modalities. The paper does not quantify how lambda affects the image/text trade-off or whether the optimal value is stable across different prompts and fabric types. Since this is a free parameter in the central mechanism, the authors should tie lambda selection to a quantitative objective (for example, CLIPScore versus structural fidelity on a validation set) and report sensitivity analysis around lambda.
minor comments (5)
  1. [Sec. 4.2.2] After Eq. (4), the sentence 'The textual features v from the text prompts T are directly extracted' reuses the symbol v for the textual feature; it should be 't' to avoid confusion with the visual feature v.
  2. [Sec. 4.2.3, Eqs. (5)-(6)] The text states that s ranges over (-1,1), but cosine similarity ranges over [-1,1]; also, with the sigmoid, alpha does not range over (0,1) as claimed, but over (lambda + (1-lambda)*sigma(-1), lambda + (1-lambda)*sigma(1)), which for lambda=0.6 is approximately (0.708, 0.892). Please correct the stated ranges.
  3. [Supp. Sec. 7.2, Table 4] The symbols 'D' and 'X' in Table 4 are not defined in the caption or the surrounding text; please add a legend explaining that 'D' denotes availability of the corresponding annotation type.
  4. [Sec. 5.4 and Supp. Sec. 9] The MLLM and user studies evaluate only 28 pairs; please state how these pairs were sampled from the 1,000 test pairs and whether the reported scores are stable across different random subsets.
  5. [Sec. 2.3] The text says 'DressCode [9] has insufficient textual content,' but reference [9] is the text-guided garment generation paper by He et al., while the virtual try-on dataset Dress Code is reference [21] (Morelli et al.); please cite the correct dataset or clarify the distinction.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical benchmark claims rest on a held-out test set, and the under-specification of the HCA mechanism is a correctness concern, not a circular one.

full rationale

HiGarment is an empirical image-generation system whose central claims are quantitative comparisons on a held-out test set of the newly collected MMDGarment dataset, together with human and MLLM evaluations. No parameter is fitted to the test labels: the trainable components (linear projection, Q-Former, and attention projections) are trained with an MSE loss on training pairs, and the only hand-set scalar, lambda=0.6 in Eq. 6, is a disclosed hyperparameter chosen from a qualitative parameter sweep, not a predicted-but-fitted constant. The HCA equations (5)-(10) represent a proposed modeling choice; their interpretation as dynamically balancing image-biased versus text-biased generation may be under-specified, especially regarding whether z replaces or supplements the base Stable Diffusion text cross-attention, but under-specification is not circularity. There are no load-bearing self-citations, no imported uniqueness theorems, and no known result merely renamed; the dataset and code are released, enabling independent external evaluation. The paper's limitations section candidly notes remaining evaluation constraints, which further supports a non-circular reading.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The ledger contains one hand-tuned numerical parameter (lambda), several standard assumptions about pretrained models, and a domain assumption about CLIP similarity reflecting modality harmony. The HCA mechanism implicitly requires an extra architectural pathway (base text cross-attention) that is not explicitly described, making it an unstated assumption that the reported equations alone do not justify.

free parameters (1)
  • lambda (lambda) in HCA weight = 0.6
    Set to 0.6 to restrict alpha to [0.6, 1] based on visual inspection of generated images; lower values lose flat-sketch details (Sec 5.5, Eq. 6). This is a hand-tuned hyperparameter, not derived from theory.
assumptions (4)
  • standard math Pre-trained Stable Diffusion v1.5, OpenCLIP ViT-H/14, and CLIP text encoder are treated as reliable frozen backbones for the task.
    The method builds directly on these pretrained models without verifying their behavior on flat garment sketches. This is standard in the field.
  • domain assumption Cosine similarity between CLIP visual and text features is a meaningful indicator of modality agreement for the FS2RG task.
    Eq. 5 uses this similarity to set the alpha balance. No evidence is provided that CLIP features are calibrated for garment-specific attributes like fabric type.
  • domain assumption The fabric database of 150 samples, plus NER and a hand-built dictionary, can map any relevant text fabric term to a useful visual sample.
    Sec 4.2.2 relies on this retrieval to enhance fabric details. The dataset contains only 11 fabric types, which may not cover the open set of user prompts.
  • ad hoc to paper The base Stable Diffusion text cross-attention pathway remains active alongside the injected HCA branch, so that scaling alpha modulates the image/text balance.
    This separation is not stated in the paper, but it is needed for Eq. 10 to produce image-biased versus text-biased behavior as claimed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HiGarment: Cross-modal Harmony Based Diffusion Model for Flat Sketch to Realistic Garment Image." pith.science (2026). https://pith.science/paper/KIBIUS6G

@misc{pith2026250523186,
  author       = {Pith},
  title        = {Pith review of: HiGarment: Cross-modal Harmony Based Diffusion Model for Flat Sketch to Realistic Garment Image},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KIBIUS6G}},
  note         = {Machine review of arXiv:2505.23186}
}
read the original abstract

Diffusion-based garment synthesis tasks primarily focus on the design phase in the fashion domain, while the garment production process remains largely underexplored. To bridge this gap, we introduce a new task: Flat Sketch to Realistic Garment Image (FS2RG), which generates realistic garment images by integrating flat sketches and textual guidance. FS2RG presents two key challenges: 1) fabric characteristics are solely guided by textual prompts, providing insufficient visual supervision for diffusion-based models, which limits their ability to capture fine-grained fabric details; 2) flat sketches and textual guidance may provide conflicting information, requiring the model to selectively preserve or modify garment attributes while maintaining structural coherence. To tackle this task, we propose HiGarment, a novel framework that comprises two core components: i) a multi-modal semantic enhancement mechanism that enhances fabric representation across textual and visual modalities, and ii) a harmonized cross-attention mechanism that dynamically balances information from flat sketches and text prompts, allowing controllable synthesis by generating either sketch-aligned (image-biased) or text-guided (text-biased) outputs. Furthermore, we collect Multi-modal Detailed Garment, the largest open-source dataset for garment generation. Experimental results and user studies demonstrate the effectiveness of HiGarment in garment synthesis. The code and dataset are available at https://github.com/Maple498/HiGarment.

Figures

Figures reproduced from arXiv: 2505.23186 by the authors.

Figure 1
Figure 1. Flat Sketch to Realistic Garment task generates a realistic sample garment from a flat sketch and text prompts, providing a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of MMDGarment dataset collection. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The overall framework of HiGarment. We design a multi-modal semantic enhancement mechanism to extract representations [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Illustration of the retrieval process in the fabric database. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of real garment image generation ability between our method and others. We use [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Ablation study results. α = 1 α = 0.9 α = 0.85 α = 0.8 α = 0.75 α = 0.7 α = 0.65 α = 0.6 α = 0.5 α = 0.4 α = 0.3 α = 0.2 Flat Sketch Text Prompt Generation Results with Different Values of α {α | 0 < α ≤ 1 } White shirt [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visual results of parameter analysis. calized details from the flat sketch (patterns on the collar or button types) begin to diminish. Therefore, we set the pa￾rameter of the Sigmoid function λ as 0.6 to restrict α ranges from 0.6 to 1 to keep a fidelity and accurate p…
Figure 8
Figure 8. Figure 8: Samples of the flat sketches and corresponding real garment images in the collected MMDGarment dataset. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Samples of the detailed description and close-ups in the collected MMDGarment dataset. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Cross-attention visualization for generated garment im [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 12
Figure 12. Figure 12: Failure cases for tiny logoes. 10. Limitations and failure cases We analyze the limitations and failure cases in this sec￾tion. Current evaluation remains limited to design-stage synthesis due to scarce public datasets and garment genera￾tion codes. We will expand com…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 34 canonical work pages

  1. [1]

    Designing for circular fashion: inte- grating upcycling into conventional garment manufacturing processes

    Reet Aus, Harri Moora, Markus Vihma, Reimo Unt, Marko Kiisa, and Sneha Kapur. Designing for circular fashion: inte- grating upcycling into conventional garment manufacturing processes. Fashion and Textiles, 8:1–18, 2021. 2

    work page 2021

  2. [2]

    Mllm-as-a-judge: Assessing multimodal llm-as-a-judge with vision-language benchmark

    Dongping Chen, Ruoxi Chen, Shilin Zhang, Yinuo Liu, Yaochen Wang, Huichi Zhou, Qihui Zhang, Pan Zhou, Yao Wan, and Lichao Sun. Mllm-as-a-judge: Assessing multimodal llm-as-a-judge with vision-language benchmark. arXiv preprint arXiv:2402.04788, 2024. 7

    arXiv 2024

  3. [3]

    Foldgen: Multimodal transformer for garment sketch- to-photo generation

    Jia Chen, Yanfang Wen, Jin Huang, Xinrong Hu, and Tao Peng. Foldgen: Multimodal transformer for garment sketch- to-photo generation. In Computer Graphics International , pages 455–466, 2023. 2, 3

    work page 2023

  4. [4]

    Viton-hd: High-resolution virtual try-on via misalignment-aware normalization

    Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In CVPR, pages 14131– 14140, 2021. 1

    work page 2021

  5. [5]

    Diffusion mod- els beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion mod- els beat gans on image synthesis. NeurIPS, 34:8780–8794,

  6. [6]

    Enhancing Label-efficient Medical Image Segmentation with Text-guided Diffusion Models

    Chun-Mei Feng. Enhancing label-efficient medical im- age segmentation with text-guided diffusion models. arXiv preprint arXiv:2407.05323, 2024. 3

    work page Pith review arXiv 2024

  7. [7]

    Cliptrans: trans- ferring visual knowledge with pre-trained models for mul- timodal machine translation

    Devaansh Gupta, Siddhant Kharbanda, Jiawei Zhou, Wan- hua Li, Hanspeter Pfister, and Donglai Wei. Cliptrans: trans- ferring visual knowledge with pre-trained models for mul- timodal machine translation. In ICCV, pages 2875–2886,

    work page

  8. [8]

    Viton: An image-based virtual try-on network

    Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In CVPR, pages 7543–7552, 2018. 3, 8

    work page 2018

Show all 42 references
  1. [9]

    Dresscode: Autoregressively sewing and gener- ating garments from text guidance

    Kai He, Kaixin Yao, Qixuan Zhang, Jingyi Yu, Lingjie Liu, and Lan Xu. Dresscode: Autoregressively sewing and gener- ating garments from text guidance. ACM Trans. Graph., 43 (4):1–13, 2024. 2

    2024

  2. [10]

    Clipscore: A reference-free evaluation met- ric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation met- ric for image captioning. In EMNLP, pages 7514–7528,

  3. [11]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. NeurIPS, 30:6629–6640, 2017. 6, 8

    2017

  4. [12]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. NeurIPS, 33:6840–6851, 2020. 3

    2020

  5. [13]

    Maskdiffusion: Exploiting pre-trained diffusion models for semantic seg- mentation

    Yasufumi Kawano and Yoshimitsu Aoki. Maskdiffusion: Exploiting pre-trained diffusion models for semantic seg- mentation. arXiv preprint arXiv:2403.11194, 2024. 3

    2024 arXiv

  6. [14]

    Blip-diffusion: Pre- trained subject representation for controllable text-to-image generation and editing

    Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pre- trained subject representation for controllable text-to-image generation and editing. NeurIPS, 36, 2024. 3

    2024

  7. [15]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 5

    2023

  8. [16]

    Ds-transunet: Dual swin transformer u-net for medical image segmentation

    Ailiang Lin, Bingzhi Chen, Jiayu Xu, Zheng Zhang, Guang- ming Lu, and David Zhang. Ds-transunet: Dual swin transformer u-net for medical image segmentation. IEEE Transactions on Instrumentation and Measurement , 71:1– 15, 2022. 3

    2022

  9. [17]

    Ed-t2v: An efficient training framework for diffusion-based text-to-video generation

    Jiawei Liu, Weining Wang, Wei Liu, Qian He, and Jing Liu. Ed-t2v: An efficient training framework for diffusion-based text-to-video generation. In IJCNN, pages 1–8, 2023. 3

    2023

  10. [18]

    Residual denoising diffu- sion models

    Jiawei Liu, Qiang Wang, Huijie Fan, Yinong Wang, Yan- dong Tang, and Liangqiong Qu. Residual denoising diffu- sion models. In CVPR, pages 2773–2783, 2024. 3

    2024

  11. [19]

    Decoupled weight decay regularization

    I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6

    2017 arXiv

  12. [20]

    U-mamba: Enhancing long-range dependency for biomedical image segmentation

    Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3

    2024 arXiv

  13. [21]

    Dress code: High- resolution multi-category virtual try-on

    Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High- resolution multi-category virtual try-on. In CVPR, pages 2231–2235, 2022. 3, 1

    2022

  14. [22]

    T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models

    Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296–4304, 2024. 2, 3

    2024

  15. [23]

    Dead- iff: An efficient stylization diffusion model with disentan- gled representations

    Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Ji- awei Liu, Lang Chen, Qian He, and Yongdong Zhang. Dead- iff: An efficient stylization diffusion model with disentan- gled representations. In CVPR, pages 8693–8702, 2024. 6, 7

    2024

  16. [24]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In ICML, pages 8748–8763, 2021. 3, 4, 6

    2021

  17. [25]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, pages 10684– 10695, 2022. 6

    2022

  18. [26]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. InECCV, pages 87–103, 2025. 3

    2025

  19. [27]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256– 2265, 2015. 3

    2015

  20. [28]

    Shape de- formation using skeleton correspondences for realistic posed fashion flat creation

    Xianmei Wan, Pik Yin Mok, and Xiaogang Jin. Shape de- formation using skeleton correspondences for realistic posed fashion flat creation. IEEE Transactions on Automation Sci- ence and Engineering, 11(2):409–420, 2014. 1

    2014

  21. [29]

    Versatile diffusion: Text, images and variations all in one diffusion model

    Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In ICCV, pages 7754– 7765, 2023. 6, 7

    2023

  22. [30]

    Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models. arXiv preprint arXiv:2308.06721,

    arXiv

  23. [31]

    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, pages 3836–3847, 2023. 3

    2023

  24. [32]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586–595,

  25. [33]

    Garmen- taligner: Text-to-garment generation via retrieval-augmented multi-level corrections

    Shiyue Zhang, Zheng Chong, Xujie Zhang, Hanhui Li, Yuhao Cheng, Yiqiang Yan, and Xiaodan Liang. Garmen- taligner: Text-to-garment generation via retrieval-augmented multi-level corrections. arXiv preprint arXiv:2408.12352 ,

    arXiv

  26. [34]

    Armani: Part-level garment-text alignment for unified cross-modal fashion design

    Xujie Zhang, Yu Sha, Michael C Kampffmeyer, Zhenyu Xie, Zequn Jie, Chengwen Huang, Jianqing Peng, and Xiaodan Liang. Armani: Part-level garment-text alignment for unified cross-modal fashion design. In ACM MM, pages 4525–4535,

  27. [35]

    Diffcloth: Diffusion based gar- ment synthesis and manipulation via structural cross-modal semantic alignment

    Xujie Zhang, Binbin Yang, Michael C Kampffmeyer, Wen- qing Zhang, Shiyue Zhang, Guansong Lu, Liang Lin, Hang Xu, and Xiaodan Liang. Diffcloth: Diffusion based gar- ment synthesis and manipulation via structural cross-modal semantic alignment. In ICCV, pages 23154–23163, 2023. 2, 3, 4

    2023

  28. [36]

    Ssr-encoder: Encoding selective subject representation for subject-driven generation

    Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jin- peng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR, pages 8069–8078,

  29. [37]

    Texcon- trol: Sketch-based two-stage fashion image generation using diffusion model

    Yongming Zhang, Tianyu Zhang, and Haoran Xie. Texcon- trol: Sketch-based two-stage fashion image generation using diffusion model. arXiv preprint arXiv:2405.04675, 2024. 2, 3

    2024 arXiv

  30. [38]

    Uni-controlnet: All-in-one control to text-to-image diffusion models

    Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 36, 2024. 3, 6, 7 HiGarment: Cross-modal Harmony Based Diffusion Model for Flat Sketch to Realistic...

    2024

  31. [39]

    Dataset construction The dataset collection team comprised six members, each assigned to three to five garment brand websites

    Dataset details 7.1. Dataset construction The dataset collection team comprised six members, each assigned to three to five garment brand websites. Images were collected from multiple garment brand websites, with specific requirements to ensure clear display of fabric and colo...

  32. [40]

    Visualization We conducted a series of visualization experiments on at- tention heat maps to demonstrate the effectiveness of our method in accurately generating garment components and attributes. Fig. 10 shows that the generated garment im- ages capture attribute details, suc...

  33. [41]

    MLLM-based evaluation and user study The MLLM evaluation contains four aspects: structure (35%), color (25%), fabric (25%), and details (15%). Specifically, structure refers to the overall layout and key components of the garment, such as the shape, style, and ar- rangement of...

  34. [42]

    Current evaluation remains limited to design-stage synthesis due to scarce public datasets and garment genera- tion codes

    Limitations and failure cases We analyze the limitations and failure cases in this sec- tion. Current evaluation remains limited to design-stage synthesis due to scarce public datasets and garment genera- tion codes. We will expand comparisons when resources [3, 33, 35] permit...

Pith tools

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