Pith. sign in

REVIEW 3 major objections 5 minor 46 references

DreamFit: Garment-Centric Human Generation via a Lightweight Anything-Dressing Encoder

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read DreamFit claims that one frozen diffusion UNet, with LoRA layers switched on only while reading the reference garment, generates texture-consistent humans from any garment at 83.4M trainable parameters, beating full-finetune baselines on…

desk verdict Genuinely novel LoRA-gated anything-dressing encoder, but the SOTA claim is undercut by a prompt-enrichment confound and thin benchmarks. read the letter →

arxiv 2412.17644 v3 pith:LGG7CLUG submitted 2024-12-23 cs.CV

Ente Lin , Xujie Zhang , Fuwei Zhao , Yuxuan Luo , Xin Dong , Long Zeng , Xiaodan Liang This is my paper · ORCID
classification cs.CV
keywords garment-centrichumangenerationdiffusionmodelsLoRAadaptiveattentionvirtualtry-ontext-to-imagelargemultimodalplug-and-playcontrol
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

DreamFit takes aim at the trade-off that splits garment-centric human generation: lightweight adapters tend to lose fine garment detail, while methods that fine-tune a second full UNet preserve texture but pay a heavy training cost and can degrade the pretrained model's priors. The paper's proposal is to let one frozen denoising UNet do both jobs. LoRA layers on its linear and convolutional layers are switched on when the reference garment image passes through and switched off when the noisy latent passes through, effectively turning the UNet into a lightweight Anything-Dressing Encoder whose features are injected through a trainable adaptive-attention branch. With 83.4M trainable parameters, DreamFit reports higher CLIP-I and CLIP-T scores than StableGarment, MagicClothing, IMAGDress, and IP-Adapter on the paper's 768 × 512 open and internal benchmarks, and it claims seamless integration with community control plugins such as ControlNet and FaceID. A large multimodal model rewrites the user's short text into a detailed garment description at inference, which the paper identifies as closing the gap between training captions and user prompts.

What carries the argument

The central object is the Anything-Dressing Encoder: the frozen denoising UNet with LoRA layers inserted into its linear and convolutional layers and gated by $I(x)$, which is 1 for the clean reference feature $c_i$ and 0 for the noisy latent $z_t$. When LoRA is on, the UNet extracts garment details; when LoRA is off, it reverts to the standard denoiser, so the same weights serve both roles. Those reference features are injected through adaptive attention, which adds a second softmax branch with trainable projections $W'_k$ and $W'_v$ (initialized from the frozen $W_k$ and $W_v$) to the normal self-attention output. Only the LoRA layers and these two projections are trained, and a pretrained large multimodal model rewrites the user's text into a detailed garment caption before inference, which the paper treats as part of the pipeline.

What would settle it

Run all four baselines and DreamFit on the same 200-image open benchmark with identical CogVLM-rewritten prompts, resolution, seed, and guidance scale, then re-score with CLIP-I, CLIP-T, aesthetic score, and human raters; if any baseline ties or exceeds DreamFit, the claimed advantage is not due to the LoRA/adaptive-attention encoder alone.

Watch

Extended reading notes

Core claim

The central claim is that a pretrained diffusion UNet already has enough visual knowledge to serve as a high-quality garment feature extractor, so no separate ReferenceNet or full UNet copy is needed. Using LoRA layers gated by $I(x)$, the same UNet behaves as the Anything-Dressing Encoder when it sees the clean reference feature $c_i$ and as the ordinary denoiser when it sees the noisy latent $z_t$. The extracted features enter generation through adaptive attention, where trainable projections $W'_k$ and $W'_v$, initialized from the frozen $W_k$ and $W_v$, let the reference garment attend to the noisy latent alongside the self-attention output. The paper reports that this configuration, combined with LMM-rewritten prompts, yields the best texture consistency, text consistency, and human-aligned preference among compared methods at 83.4M trainable parameters, roughly one-tenth the full-finetuning budget.

Load-bearing premise

The superiority claim rests on two 200-image benchmarks, one handpicked from public test sets and one gathered from the internet, judged by automatic similarity scores, an aesthetic scorer, and a large language model; if those selections are not representative, or if the richer text prompts used by DreamFit give it an advantage the baselines do not receive, the reported margins could change.

Editorial extensions

If this is right

  • If DreamFit is right, the standard practice of copying the whole UNet to encode garments is unnecessary: the same frozen backbone can switch between reading the garment and denoising, cutting trainable parameters from roughly 875M to 83.4M.
  • Because the base diffusion weights stay frozen, the method should inherit the generalization of the pretrained model, which the paper offers as the reason it handles non-garment references, creative styles, and varied prompt instructions.
  • The added modules sit beside the frozen UNet, so community control plugins such as ControlNet and FaceID can be combined with DreamFit without retraining, making it a drop-in layer for existing workflows.
  • Prompt rewriting is a measured part of the method: ablations show CLIP-I, aesthetic score, and human preference all improve when simple user prompts are replaced by LMM-generated detailed garment descriptions, so closing the train/inference prompt gap is one of the claimed levers.
  • The recipe transfers to newer backbones: in the paper's supplementary results, SDXL and FLUX versions improve on the SD1.5 version, with FLUX best at preserving fine-grained logos and textures.

Reading between the lines

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

  • The LoRA gate applies to any clean reference image, not only garments, so the same architecture is a candidate generic reference-conditioning layer for identity, object, or style customization, which the paper names as future work.
  • A fair comparison that feeds every baseline the same LMM-rewritten prompts would separate the encoder's contribution from prompt enrichment; the paper's ablations measure the prompt effect only for DreamFit itself.
  • Because the parameter count is largely independent of the base model, the training recipe could be replayed quickly on each new diffusion backbone, and the paper's SDXL and FLUX experiments already hint at that scalability.
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 proposes DreamFit, a garment-centric human generation method built on a frozen Stable Diffusion UNet with trainable LoRA layers and two adaptive-attention projection layers. During training, LoRA layers are switched on only for the clean reference image, turning the UNet into a lightweight 'Anything-Dressing Encoder'; at inference, the user prompt is rewritten by CogVLM to reduce the training-inference prompt gap. The authors report that DreamFit outperforms IP-Adapter, StableGarment, MagicClothing, and IMAGDress on CLIP-T, CLIP-I, Aesthetic Score, and a GPT-4o-based Human-Aligned Score on two 200-image benchmarks at 768x512, while using 83.4M trainable parameters versus roughly 875M for full fine-tuning. The paper also demonstrates integration with ControlNet and FaceID and reports ablations of the LoRA/Adapter components and of the prompt-rewriting step.

Significance. If the quantitative claims are established, the paper makes a useful contribution: it shows that a frozen diffusion UNet can be repurposed as a reference-image encoder through LoRA gating, avoiding a bulky ReferenceNet, and that LMM-based prompt enrichment can improve generation fidelity. The reported parameter count is substantially lower than full fine-tuning, and the plug-and-play demonstrations with ControlNet and FaceID are appealing for practitioners. However, the headline 'state-of-the-art' claim currently rests on a comparison that may be confounded by an unstated prompt-rewriting asymmetry, on single-run metric values without uncertainty, and on a GPT-4o-based preference score rather than a human preference study. These issues are load-bearing for the central claim, so the contribution is not yet fully validated.

major comments (3)
  1. [Methodology, 'Training and inference'; Experiment, 'Quantitative Results', Table 1] The SOTA comparison in Table 1 is confounded by the LMM prompt-rewriting step. The paper states that during inference DreamFit rewrites user text with CogVLM (Section 'Training and inference'), and the ablation 'Impact of Text Prompts' (Table 3) shows that rewriting alone raises CLIP-I from 0.663 (simple text) to 0.687 (Ours) and the human-study preference from 0.10 to 0.87. Table 1 does not state whether StableGarment, MagicClothing, IMAGDress, and IP-Adapter received the same LMM-refined prompts. The open-benchmark description says each garment is associated with refined prompts, which may control the confound there, but the internal-benchmark description does not say this, and DreamFit's own inference pipeline again applies CogVLM. Since the internal-benchmark margin (CLIP-I 0.687 vs 0.629 for StableGarment) is comparable in size to the measured prompt-rewriting effect, the reported superiority may be largely due to prompt enrichment rather than to the proposed Anything-Dressing Encoder. Please clarify whether baselines received identical rewritten prompts in each benchmark, and if not, rerun the comparison under matched prompt conditions or explicitly report both settings.
  2. [Experiment, 'Quantitative Results', Table 1] Table 1 reports single-run point estimates without error bars, confidence intervals, or significance tests. On 200-image benchmarks, differences such as CLIP-I 0.740 vs 0.720 (open benchmark) and CLIP-T 0.334 vs 0.320 (internal benchmark) may be within sampling noise; the absence of uncertainty estimates is especially problematic given that the prompt-rewriting ablation in Table 3 produces shifts of 0.02-0.03 CLIP-I. Additionally, the 'Human-Aligned Score' is a GPT-4o preference score, not a human preference study; the only human study in the paper (Table 3) evaluates text-prompt configurations, not method superiority. Please provide repeated evaluations or bootstrap intervals for all metrics and either add a human preference study for the method comparison or relabel HAS as an automated model-based score.
  3. [Experiment, 'Datasets'] The two benchmarks are 200 images each, and the open benchmark is described as 'handpicked' by the authors while the internal benchmark is 'manually gathered' from the internet. No selection protocol, inclusion criteria, or release plan is given. This creates a risk of selection bias favoring the authors' method and raises reproducibility concerns. Please specify how the 200 garments were sampled (e.g., random stratified sampling versus cherry-picking), report the exact source splits from VITON-HD/DressCode, and make the benchmark lists available or provide a detailed deterministic protocol so others can reconstruct the evaluation.
minor comments (5)
  1. [Throughout] There are numerous typos and grammatical errors, e.g., 'the weights of our denoising UNet the frozen' (Methodology, 'Adaptive attention'), 'Language Model Models (LMMs)' (Methodology, 'Training and inference'), 'V AE' (Preliminary), 'A image of person' (Figure 6), and inconsistent capitalization of 'DreamFit'/'Dreamfit'. A careful proofread is needed.
  2. [Experiment, 'Quantitative Results'] The sentence 'The quantitative results are reported in Tab 1. ... on three metrics' is inaccurate: Table 1 reports four metrics (CLIP-T, CLIP-I, AS, HAS). Please correct the count.
  3. [Methodology, Eq. (3)] In Eq. (3), the notation Delta W(x) should be defined more precisely: Delta W is typically a low-rank matrix applied to x, so writing Delta W (x) without parentheses or with an explicit application would improve clarity.
  4. [Experiment, 'Baselines'] The comparison includes only four baselines, all based on SD1.5. The abstract's 'surpasses all existing methods' is stronger than what is demonstrated; please restrict the claim to the evaluated baselines or add more recent methods such as OOTDiffusion and TryOnAdapter.
  5. [Supplementary Material, 'Limitations and future work'] The stated OOM limitation for the FLUX version at 1024x768 is interesting but should be mentioned in the main text as a limitation, since the abstract emphasizes lightweight training and scalability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DreamFit's contribution is an architectural combination trained with the standard diffusion objective; the SOTA claim is an external benchmark comparison, and the mild prompt-enrichment confound is an evaluation gap, not a derivation that reduces to its inputs.

full rationale

DreamFit's stated derivation chain is an architecture plus a standard diffusion objective, not a parameter-free derivation that could collapse into its own inputs. The Anything-Dressing Encoder is defined by LoRA gates I(x)=0/1 and adaptive attention z_new = Softmax(QK^T/sqrt(d))V + Softmax(Q(K')^T/sqrt(d))V'; these equations are constructional, not circular, because the encoder is literally the UNet with LoRA activated and the adaptive attention is an injected adapter. The training objective LLDM is the usual denoising loss, so no fitted constant is renamed as a prediction. The SOTA claim is an empirical comparison against external released baselines on two hand-built benchmarks; no equation in the paper equates a claimed output with an input, and no self-citation is load-bearing. The only notable concern is that Table 1 does not state whether baselines also received the LMM-enriched prompts that DreamFit uses at inference, and CLIP-T is computed against the prompt condition; however, the paper openly identifies prompt rewriting as a component and ablates it in Table 3, so this is an evaluation-protocol confound, not a circular reduction. Therefore no specific circular step can be exhibited, and the appropriate finding is no significant circularity.

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

The paper does not introduce new physical entities or forces. It introduces an architectural module, the Anything-Dressing Encoder, built from existing components (LoRA, adapters, attention). The main free parameters are training hyperparameters and the choice of LMM prompt rewriter, none of which are fitted to a target result in a predictive sense.

free parameters (5)
  • LoRA rank = 64 (SD1.5, SDXL); 32 (FLUX)
    Hyperparameter chosen by hand; controls capacity of the anything-dressing encoder.
  • Guidance scale w = 7.5
    Classifier-free guidance weight at inference, chosen without reported sweep.
  • Learning rate = 1e-4
    AdamW learning rate for all training runs.
  • Training steps = 90k (SD1.5, SDXL); 20k (FLUX)
    Training duration chosen by hand.
  • LMM prompt rewriter = CogVLM
    Inference pipeline uses CogVLM to rewrite prompts; choice of LMM not ablated.
assumptions (5)
  • domain assumption A frozen Stable Diffusion UNet can serve as a feature extractor for clean reference images after adding trainable LoRA layers.
    Core architectural premise of the Anything-Dressing Encoder; validated only by the reported results, not by analysis.
  • domain assumption Pretrained diffusion priors are preserved when denoising with LoRA turned off.
    The gate function I(x) assumes reference encoding and denoising can share weights without interference.
  • domain assumption LMM-generated captions reduce the train/inference prompt gap and improve fidelity.
    The paper uses LMMs at inference but not at training; this asymmetry is assumed beneficial.
  • domain assumption CLIP-I, CLIP-T, Aesthetic Score, and GPT-4o HAS are valid proxies for garment consistency and human preference.
    Metrics used for all comparisons; no calibration against human judgments for the main benchmark tables.
  • domain assumption The 200-image open and 200-image internal benchmarks are representative and unbiased.
    The open benchmark items are handpicked by the authors; the internal benchmark is author-collected from the internet.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DreamFit: Garment-Centric Human Generation via a Lightweight Anything-Dressing Encoder." pith.science (2026). https://pith.science/paper/LGG7CLUG

@misc{pith2026241217644,
  author       = {Pith},
  title        = {Pith review of: DreamFit: Garment-Centric Human Generation via a Lightweight Anything-Dressing Encoder},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LGG7CLUG}},
  note         = {Machine review of arXiv:2412.17644}
}
abstract

Diffusion models for garment-centric human generation from text or image prompts have garnered emerging attention for their great application potential. However, existing methods often face a dilemma: lightweight approaches, such as adapters, are prone to generate inconsistent textures; while finetune-based methods involve high training costs and struggle to maintain the generalization capabilities of pretrained diffusion models, limiting their performance across diverse scenarios. To address these challenges, we propose DreamFit, which incorporates a lightweight Anything-Dressing Encoder specifically tailored for the garment-centric human generation. DreamFit has three key advantages: (1) \textbf{Lightweight training}: with the proposed adaptive attention and LoRA modules, DreamFit significantly minimizes the model complexity to 83.4M trainable parameters. (2)\textbf{Anything-Dressing}: Our model generalizes surprisingly well to a wide range of (non-)garments, creative styles, and prompt instructions, consistently delivering high-quality results across diverse scenarios. (3) \textbf{Plug-and-play}: DreamFit is engineered for smooth integration with any community control plugins for diffusion models, ensuring easy compatibility and minimizing adoption barriers. To further enhance generation quality, DreamFit leverages pretrained large multi-modal models (LMMs) to enrich the prompt with fine-grained garment descriptions, thereby reducing the prompt gap between training and inference. We conduct comprehensive experiments on both $768 \times 512$ high-resolution benchmarks and in-the-wild images. DreamFit surpasses all existing methods, highlighting its state-of-the-art capabilities of garment-centric human generation.

Figures

Figures reproduced from arXiv: 2412.17644 by the authors.

Figure 1
Figure 1. Garment-centric human generation results of our DreamFit: [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Performance comparison between baselines and [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of DreamFit. Our method constructs an Anything-Dressing Encoder utilizing LoRA layers. The reference [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on the open and internal benchmarks. DreamFit demonstrates a distinct advantage in handling [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Plug-and-play results of DreamFit, our method can seamlessly integrate with community conditional control plugins. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results of the ablation study on network [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: These results confirm DreamFit’s impressive gen [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 8
Figure 8. Figure 8: More qualitative results at 768 resolution on internal benchmark. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results of SDXL version at 1024 resolution on the Open benchmark. Please zoom in for more details. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Qualitative results of SDXL version at 1024 resolution on the open benchmark. Please zoom in for more details. [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Qualitative results of FLUX version at 1024 resolution on the open benchmark. Please zoom in for more details. [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Qualitative comparison results between SDXL and FLUX versions at 1024 resolution. FLUX version significantly [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: The Human-Aligned evaluation pipeline. We utilize GPT-4o for conducting automated evaluations. [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Visualized results of the outputs and evaluation reasons corresponding to different scores. [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: interface of the questionnaire used to evaluate the impact of text prompts. [PITH_FULL_IMAGE:figures/full_fig_p016_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 11 canonical work pages

  1. [1]

    Deepspeed

    2023. Deepspeed. Https://github.com/microsoft/DeepSpeed

    work page 2023

  2. [2]

    L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al

    Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F. L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  3. [3]

    Chen, W.; Gu, T.; Xu, Y.; and Chen, C. 2024 a . Magic Clothing: Controllable Garment-Driven Image Synthesis. arXiv preprint arXiv:2404.09512

    arXiv 2024

  4. [4]

    Chen, X.; Huang, L.; Liu, Y.; Shen, Y.; Zhao, D.; and Zhao, H. 2024 b . Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 6593--6602

    work page 2024

  5. [5]

    Choi, S.; Park, S.; Lee, M.; and Choo, J. 2021. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    work page 2021

  6. [6]

    Choi, Y.; Kwak, S.; Lee, K.; Choi, H.; and Shin, J. 2024. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139

    arXiv 2024

  7. [7]

    Cui, A.; Mahajan, J.; Shah, V.; Gomathinayagam, P.; and Lazebnik, S. 2023. Street TryOn: Learning In-the-Wild Virtual Try-On from Unpaired Person Images. arXiv preprint arXiv:2311.16094

    arXiv 2023

  8. [8]

    Han, Y.; Zhu, J.; He, K.; Chen, X.; Ge, Y.; Li, W.; Li, X.; Zhang, J.; Wang, C.; and Liu, Y. 2024. Face Adapter for Pre-Trained Diffusion Models with Fine-Grained ID and Attribute Control. arXiv preprint arXiv:2405.12970

    arXiv 2024

Show all 46 references
  1. [9]

    L.; and Choi, Y

    Hessel, J.; Holtzman, A.; Forbes, M.; Bras, R. L.; and Choi, Y. 2021. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718

    2021 arXiv

  2. [10]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems

    2020

  3. [11]

    Ho, J.; and Salimans, T. 2022. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598

    2022 arXiv

  4. [12]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

    2021 arXiv

  5. [13]

    Hu, L. 2024. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8153--8163

    2024

  6. [14]

    Kim, J.; Gu, G.; Park, M.; Park, S.; and Choo, J. 2024. StableVITON: Learning Semantic Correspondence with Latent Diffusion Model for Virtual Try-On. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2024

  7. [15]

    P.; and Welling, M

    Kingma, D. P.; and Welling, M. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114

    2013 arXiv

  8. [16]

    Ku, M.; Jiang, D.; Wei, C.; Yue, X.; and Chen, W. 2023. Viescore: Towards explainable metrics for conditional image synthesis evaluation. arXiv preprint arXiv:2312.14867

    2023 arXiv

  9. [17]

    Labs, B. F. 2024. FLUX.1-dev. Https://huggingface.co/black-forest-labs/FLUX.1-dev

    2024

  10. [18]

    Li, S.; Yang, L.; Jiang, X.; Lu, H.; Di, Z.; Lu, W.; Chen, J.; Liu, K.; Yu, Y.; Lan, T.; et al. 2024. SwiftDiffusion: Efficient Diffusion Model Serving with Add-on Modules. arXiv preprint arXiv:2407.02031

    2024 arXiv

  11. [19]

    Liu, H.; Li, C.; Li, Y.; and Lee, Y. J. 2024 a . Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 26296--26306

    2024

  12. [20]

    Liu, J.; Li, C.; Sun, Q.; Ming, J.; Fang, C.; Wang, J.; Zeng, B.; and Liu, S. 2024 b . Ada-adapter: Fast Few-shot Style Personlization of Diffusion Model with Pre-trained Image Encoder. arXiv preprint arXiv:2407.05552

    2024 arXiv

  13. [21]

    Morelli, D.; Baldrati, A.; Cartella, G.; Cornia, M.; Bertini, M.; and Cucchiara, R. 2023. LaDI-VTON: Latent Diffusion Textual-Inversion Enhanced Virtual Try-On . In Proceedings of the ACM International Conference on Multimedia

    2023

  14. [22]

    Morelli, D.; Fincato, M.; Cornia, M.; Landi, F.; Cesari, F.; and Cucchiara, R. 2022. Dress code: High-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2231--2235

    2022

  15. [23]

    Mou, C.; Wang, X.; Xie, L.; Wu, Y.; Zhang, J.; Qi, Z.; and Shan, Y. 2024. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 4296--4304

    2024

  16. [24]

    Peng, Y.; Cui, Y.; Tang, H.; Qi, Z.; Dong, R.; Bai, J.; Han, C.; Ge, Z.; Zhang, X.; and Xia, S.-T. 2024. Dreambench++: A human-aligned benchmark for personalized image generation. arXiv preprint arXiv:2406.16855

    2024 arXiv

  17. [25]

    Podell, D.; English, Z.; Lacey, K.; Blattmann, A.; Dockhorn, T.; M \"u ller, J.; Penna, J.; and Rombach, R. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952

    2023 arXiv

  18. [26]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning

    2021

  19. [27]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022 a . High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10684--10695

    2022

  20. [28]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022 b . High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    2022

  21. [29]

    Ruiz, N.; Li, Y.; Jampani, V.; Pritch, Y.; Rubinstein, M.; and Aberman, K. 2023. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 22500--22510

    2023

  22. [30]

    Schuhmann, C.; Beaumont, R.; Vencu, R.; Gordon, C.; Wightman, R.; Cherti, M.; Coombes, T.; Katta, A.; Mullis, C.; Wortsman, M.; et al. 2022. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems,...

    2022

  23. [31]

    Shen, F.; Jiang, X.; He, X.; Ye, H.; Wang, C.; Du, X.; Li, Z.; and Tang, J. 2024. IMAGDressing-v1: Customizable Virtual Dressing. arXiv preprint arXiv:2407.12705

    2024 arXiv

  24. [32]

    Song, J.; Meng, C.; and Ermon, S. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502

    2020 arXiv

  25. [33]

    Wang, H.; Zhang, Z.; Di, D.; Zhang, S.; and Zuo, W. 2024 a . MV-VTON: Multi-View Virtual Try-On with Diffusion Models. arXiv preprint arXiv:2404.17364

    2024 arXiv

  26. [34]

    Wang, R.; Guo, H.; Liu, J.; Li, H.; Zhao, H.; Tang, X.; Hu, Y.; Tang, H.; and Li, P. 2024 b . StableGarment: Garment-Centric Generation via Stable Diffusion. arXiv preprint arXiv:2403.10783

    2024 arXiv

  27. [35]

    Wang, W.; Lv, Q.; Yu, W.; Hong, W.; Qi, J.; Wang, Y.; Ji, J.; Yang, Z.; Zhao, L.; Song, X.; et al. 2023. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079

    2023 arXiv

  28. [36]

    Xing, J.; Xu, C.; Qian, Y.; Liu, Y.; Dai, G.; Sun, B.; Liu, Y.; and Wang, J. 2024. TryOn-Adapter: Efficient Fine-Grained Clothing Identity Adaptation for High-Fidelity Virtual Try-On. arXiv preprint arXiv:2404.00878

    2024 arXiv

  29. [37]

    Xu, Y.; Gu, T.; Chen, W.; and Chen, C. 2024. Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779

    2024 arXiv

  30. [38]

    xujie zhang; Li, X.; Kampffmeyer, M.; Dong, X.; Xie, Z.; Zhu, F.; Dong, H.; and Liang, X. 2023. WarpDiffusion: Efficient Diffusion Model for High-Fidelity Virtual Try-on. arXiv:2312.03667

    2023 arXiv

  31. [39]

    Yang, Y.; Wang, W.; Peng, L.; Song, C.; Chen, Y.; Li, H.; Yang, X.; Lu, Q.; Cai, D.; Wu, B.; et al. 2024. LoRA-Composer: Leveraging Low-Rank Adaptation for Multi-Concept Customization in Training-Free Diffusion Models. arXiv preprint arXiv:2403.11627

    2024 arXiv

  32. [40]

    Ye, H.; Zhang, J.; Liu, S.; Han, X.; and Yang, W. 2023. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721

    2023 arXiv

  33. [41]

    Zhang, L.; Rao, A.; and Agrawala, M. 2023. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 3836--3847

    2023

  34. [42]

    Zhang, X.; Lin, E.; Li, X.; Luo, Y.; Kampffmeyer, M.; Dong, X.; and Liang, X. 2024. MMTryon: Multi-Modal Multi-Reference Control for High-Quality Fashion Generation. arXiv preprint arXiv:2405.00448

    2024 arXiv

  35. [43]

    Zhu, L.; Li, Y.; Liu, N.; Peng, H.; Yang, D.; and Kemelmacher-Shlizerman, I. 2024. M&M VTO: Multi-Garment Virtual Try-On and Editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1346--1356

    2024

  36. [44]

    Zhu, L.; Yang, D.; Zhu, T.; Reda, F.; Chan, W.; Saharia, C.; Norouzi, M.; and Kemelmacher-Shlizerman, I. 2023. TryOnDiffusion: A Tale of Two UNets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2023

  37. [45]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  38. [46]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.