Pith. sign in

REVIEW 3 major objections 6 minor 42 references

TalkFashion: Intelligent Virtual Try-On Assistant Based on Multimodal Large Language Model

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Text-only try-on assistant beats six baselines on VITON-HD.

desk verdict Useful system paper with a genuinely new text-to-database router and mask-free local editing, but the headline numbers are clouded by an unmeasured overlap between the retrieval database and the test set. read the letter →

arxiv 2507.05790 v1 pith:GNRM32LC submitted 2025-07-08 cs.CV

Yujie Hu , Xuanyu Zhang , Weiqi Li , Jian Zhang This is my paper · ORCID
classification cs.CV
keywords virtualtry-onmultimodallargelanguagemodelsdiffusiontext-guidedimageeditinglocalrepaintingfunctioncallingimage-textmatchingmask-free
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

TalkFashion is a virtual try-on assistant that takes a user photo and a plain-text instruction and decides by itself whether to perform a full outfit change or a local edit. For full changes, it matches the instruction against an in-shop garment database; when the match score is high enough it runs an image-based try-on model with the retrieved garment, otherwise it falls back to a text-based generation model. For local edits, a multimodal LLM writes a detailed description that guides an SDXL inpainting network, while a spatial comprehension model passes a segmentation embedding to SAM so the edit region is masked automatically, with no manual input. On VITON-HD, the paper reports the best or second-best values across PSNR, SSIM, LPIPS, FID, and CLIP-Score for both full outfit change and localized editing.

What carries the argument

The central machinery is the invocation module plus the matching gate. After the LLM emits a structured response (Eq. 1), the module parses it into an item, a detail, and a function; for full outfit change, a cosine-similarity score between the text embedding and each garment image embedding in the in-shop database decides whether the image-based try-on model or the text-based one is executed (Eq. 2). For localized editing, the key object is the segmentation token embedding $h_{\langle seg\rangle}$ extracted from the spatial comprehension model and passed to SAM, which produces the mask $m$ without manual input; the MLLM's refined description is projected into the generative space as $E_{\mathrm{guide}}$ to steer the diffusion network (Eqs. 4–5). Together these components are what allow a single text instruction to route across three formerly separate try-on tasks.

What would settle it

Rerun the full outfit change experiments with all VITON-HD test garments removed from the in-shop database; if the PSNR, SSIM, and FID advantages over text-only baselines largely disappear, the reported superiority is retrieval-backed rather than generation-backed.

Watch

Extended reading notes

Core claim

The central claim is that a single text-only interface can unify image-based virtual try-on, text-based virtual try-on, and mask-free local editing, and that this unification produces better semantic consistency and visual quality than any one end-to-end approach. The function invocation is a decision rule: the LLM outputs a structured call (Eq. 1) selecting the item, the edit detail, and the function; for full outfit change, the system computes a cosine-similarity match score $MS(d,\mathcal{D})$ between the user text and garment images in the in-shop database and uses image-based try-on $G_{\mathrm{img}}$ when the score passes a threshold $\tau$, otherwise text-based try-on $G_{\mathrm{txt}}$ (Eq. 2). Localized editing (Eqs. 4–5) replaces manual masks with a mask $m$ produced by SAM from the spatial comprehension model's embedding $h_{\langle seg\rangle}$, while the MLLM supplies the semantic guide $E_{\mathrm{guide}}$. The reported experiments on VITON-HD, including an ablation of the MLLM that improves FID from 13.208 to 9.614 and CLIP-Score from 26.628 to 27.236, support the claim.

Load-bearing premise

The numerical advantage on full outfit change hinges on an unstated overlap between the in-shop clothes database and the VITON-HD test set; if the test garments are inside the database, the matching gate retrieves the ground-truth garment and the reported scores measure retrieval-backed try-on rather than text-only generation.

Editorial extensions

If this is right

  • A user can request a full outfit change or a local attribute edit in plain language, without uploading a reference garment image or drawing a mask.
  • When the requested garment is not in the in-shop database, the matching gate automatically switches from image-based to text-based generation instead of failing.
  • The mask-free segmentation route (SCM + SAM) extends fine-grained editing to fashion attributes such as sleeve length and neckline without any manual annotation at test time.
  • Because the pipeline uses off-the-shelf components (LLM, BLIP2, SAM, SDXL) with fine-tuning, each component can be swapped for a stronger model as they become available.
  • The multi-turn conversational wrapper allows follow-up requests, such as changing the color or sleevelessness after a first edit.

Reading between the lines

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

  • If the retrieval path is the main source of the reported PSNR/SSIM gains, a controlled holdout of test garments from the database would likely shrink the gap over text-only baselines; this is our editorial prediction, not a claim in the paper.
  • The same invocation-and-matching architecture could be transferred to other object-editing assistants (e.g., hairstyle or interior design) that pair free-form text with a candidate database of reference images.
  • The segmentation-token-to-SAM mechanism may generalize beyond fashion, since it couples a language model's semantic output with a general-purpose segmenter; the paper does not test this.
  • The manual annotation of ten fashion attributes limits the local editing to that vocabulary; scaling the mask dataset would be needed for open-domain instructions.
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 / 6 minor

Summary. This paper presents TalkFashion, a virtual try-on assistant that uses a large language model (Qwen2-7B-Instruct) to parse user text instructions and route them to one of three pipelines: an image-based try-on model (activated when the instruction matches an item in an in-shop clothing database with a score above a threshold), a text-based try-on model, or a novel localized editing model that employs Qwen2-VL and SAM to automatically generate an edit mask from the text instruction. The authors report quantitative results on VITON-HD for both full outfit change and localized editing, claiming superior performance over several baselines, and an ablation study on the MLLM prompt refiner.

Significance. If the evaluation were clean, the paper would offer a practical engineering contribution: an LLM-based invocation mechanism that unifies image- and text-conditioned try-on, and a mask-free localized editing pipeline that avoids manual mask annotation. The use of off-the-shelf components and the two-stage training procedure are pragmatic, and the ablation on the MLLM refiner provides some evidence for its efficacy. However, the central quantitative claim of superiority over current methods is not yet supported because the full-outfit-change evaluation may suffer from database leakage, the routing threshold is undisclosed, and the baselines are not dedicated virtual try-on methods. These issues need to be resolved before the results can be interpreted.

major comments (3)
  1. [Section IV.A, Eq. (2), Table I] The in-shop clothes database is said to "primarily consist" of VITON-HD, and the evaluation is also performed on VITON-HD. The paper never states whether the test-split garment images are excluded from the database. Under Eq. (2), if the ground-truth garment for a full-outfit-change query is present in the database, the BLIP2 matching will retrieve that exact garment and route the query to the image-based model Gimg. The reported PSNR of 41.854 and SSIM of 0.995 are consistent with near-reconstruction of the original image, which would mean the comparison in Table I measures retrieval plus conventional try-on rather than text-guided generation, and it gives Ours an advantage not available to the mask-required baselines. To support the claim of better text-guided full-outfit generation, the authors must either use a garment database disjoint from the test set, or evaluate the text-only branch Gtxt separately, and report the fraction of queries routed to each branch.
  2. [Section III.B, Eq. (2), Limitations] The routing threshold τ is a free parameter that determines how often the image-based route is used, yet its value is never disclosed. The Limitations section concedes that the "predefined matching score thresholds" depend on the choice of image-based and text-based models. Without reporting τ, the sensitivity of Table I's results to this parameter, or at least the distribution of match scores for the test queries, the reported metrics cannot be reproduced and the balancing between the image-based and text-based branches is uninterpretable. A sensitivity analysis over τ and a separate evaluation of Gtxt are needed.
  3. [Section IV.B, Table I] The full-outfit-change comparison is made against general inpainting/editing models (SDXL-Inpainting, ControlNet-Inpainting, UltraEdit, MagicQuill), not against dedicated virtual try-on methods. Recent VTON baselines such as HR-VITON, OOTDiffusion, StableVITON, or CatVTON do not appear. Since these general models are not designed to preserve garment identity or to warp clothing to the person's pose, the comparison does not substantiate the claim that TalkFashion outperforms "the current methods" in virtual try-on. The authors should add at least two recent image-conditioned try-on baselines under the same evaluation protocol.
minor comments (6)
  1. [Table II caption] The word "Abalation" in the caption should be "Ablation".
  2. [Section IV.A] The size of the in-shop clothes database and the exact split with respect to the VITON-HD test set should be stated explicitly.
  3. [Section IV.A] The statement "PSNR, SSIM, and LPIPS are used to measure the fidelity between the reconstructed image and the original image" is ambiguous; it should be clarified whether the instruction in this setup describes the original clothes, so that the goal is reconstruction.
  4. [Section IV.B] The claim that "Our method achieves optimal performance across nearly all metrics" is too broad; for example, MagicQuill achieves a lower FID than Ours in the localized editing task (8.297 vs. 9.614).
  5. [General] The paper refers to "Supplementary Material" several times, but the supplementary file is not included with the submission; the prompt template details and mask annotation examples should be summarized in the main text if no supplementary is provided.
  6. [Figure 5] The "Original" column appears both at the far left and far right of the figure; this duplication is visually confusing and should be removed or relabeled.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: TalkFashion is an assembly of externally benchmarked components, and its claims rest on empirical comparisons rather than self-referential definitions.

full rationale

No circular step is present. The claimed contribution is a system that combines existing, externally validated components (Qwen2, BLIP2, SAM, DensePose, SDXL, VITON-HD, GPT-4o) with a matching-based routing rule in Eq. (2) and a mask-free local-editing pipeline in Eqs. (4)-(5). Neither equation defines an output in terms of the metric used to evaluate it, nor is any predicted quantity fitted from the test target. The authors' own prior works appear only as background references (EditGuard, T2I-Adapters, DragDiffusion, DiffEditor) and are not used to justify the paper's central architecture or results. The ablation in Table II is an honest comparison against an attribute-tag baseline. The closest caveat, that the in-shop clothes database "primarily consists of" VITON-HD while Table I is evaluated on VITON-HD, is a potential benchmark-overlap issue for the full-outfit scores, but it is an experimental validity concern rather than a circular derivation: the paper never claims a first-principles result, and the routing rule in Eq. (2) is not self-referential. The Limitations section also acknowledges threshold sensitivity, which further indicates that the matching threshold is a tunable component rather than a hidden redefinition of the evaluation target. Therefore, the derivation chain is self-contained and non-circular.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on four domain assumptions and two hand-set choices, listed above. No new physical or formal entities are introduced; every module (Qwen2, Qwen2-VL, BLIP2, SAM, SDXL, DensePose, human parsing) is an off-the-shelf model from prior work. The free parameters that constrain the claim are the unreported routing threshold tau and the hand-chosen 10-area annotation taxonomy defining the editable attribute space.

free parameters (2)
  • routing threshold tau = not reported
    Gates image-based vs. text-based try-on in Eq. (2). Its value is never given, and the Limitations section says it depends on the chosen try-on models, so it is a hand-set constant whose influence on the reported scores is unknown.
  • annotation taxonomy (10 garment areas) = 10 categories (sleeve length, neckline, length, etc.)
    Defines the editable attribute space for localized editing (Section IV.A). The 'fully automated' editing claim holds only within this hand-chosen taxonomy, which is never validated against real user requests.
assumptions (4)
  • domain assumption BLIP2 cosine similarity between user text and garment images is a valid proxy for user intent
    Invoked in Section III.B as the match score MS(d, D) that routes between Gimg and Gtxt. The paper provides no evidence that this similarity corresponds to user preferences, and the threshold on it is unreported.
  • domain assumption The 5,000 GPT-4o-generated instruction-mask pairs cover real user editing requests
    The spatial comprehension model is trained on these pairs (Section IV.A), and the localized editing evaluation uses fixed-format instructions of the same style (Fig. 5), so test instructions match the training distribution by construction.
  • domain assumption The in-shop database and the VITON-HD test set can overlap without invalidating the full outfit change scores
    Section IV.A builds the database from VITON-HD and Section IV.B evaluates on VITON-HD; the paper neither reports the overlap nor argues it is harmless.
  • domain assumption Qwen2-7B-Instruct invokes the correct function, clothing item, and edit details for arbitrary instructions
    Eq. (1) treats {F, i, d} = LLM(xtxt, P) as correct, but no accuracy metric for this invocation step is reported anywhere in Section IV.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TalkFashion: Intelligent Virtual Try-On Assistant Based on Multimodal Large Language Model." pith.science (2026). https://pith.science/paper/GNRM32LC

@misc{pith2026250705790,
  author       = {Pith},
  title        = {Pith review of: TalkFashion: Intelligent Virtual Try-On Assistant Based on Multimodal Large Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GNRM32LC}},
  note         = {Machine review of arXiv:2507.05790}
}
read the original abstract

Virtual try-on has made significant progress in recent years. This paper addresses how to achieve multifunctional virtual try-on guided solely by text instructions, including full outfit change and local editing. Previous methods primarily relied on end-to-end networks to perform single try-on tasks, lacking versatility and flexibility. We propose TalkFashion, an intelligent try-on assistant that leverages the powerful comprehension capabilities of large language models to analyze user instructions and determine which task to execute, thereby activating different processing pipelines accordingly. Additionally, we introduce an instruction-based local repainting model that eliminates the need for users to manually provide masks. With the help of multi-modal models, this approach achieves fully automated local editings, enhancing the flexibility of editing tasks. The experimental results demonstrate better semantic consistency and visual quality compared to the current methods.

Figures

Figures reproduced from arXiv: 2507.05790 by the authors.

Figure 1
Figure 1. Proposed intelligent virtual try-on assistant named TalkFashion. Users can achieve various outfits by simply inputting text instructions. paints over the area to be changed and then injects the garment information to generate a complete piece of clothing that matches the reference image. This approach does not consider the user’s finer-grained needs, such as modifying only a part of a garment while preserving the ov… view at source ↗
Figure 2
Figure 2. Overview of the TalkFashion framework. We first combine user-input text instructions xtxt with a prompt template P and input into the LLM➀ to generate structured responses➁, detailing the functions to be used along with other useful information. Then segment the area to be modified according to the selected function F➂. If the selected function is full outfit change, the user’s instruction is matched with items in t… view at source ↗
Figure 3
Figure 3. The pipeline of localized editing. During the inference process, the image ximg and text instruction xtxt are fed into the multi-modal large language model to obtain a detailed prompt. This prompt is first mapped into a generative space and provides semantic guidance through cross-attention mechanisms. Simultaneously, ximg and xtxt are processed by the spatial comprehension model, and the last-layer embedding for th… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons on full outfit change. Our method is superior to other methods in generating realistic images and ensuring semantic consistency on VITON-HD. model MLLM for semantic comprehension. On one side, the text instruction xtxt and input image ximg are i…
Figure 5
Figure 5. Figure 5: Qualitative comparisons on localized editing. We see that the editing effect of other methods is not obvious, and it is easy to make mistakes in color and detail. Our method understands the instructions well and guarantees the fidelity. TABLE II: Abalation on MLLM. We …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 31 canonical work pages

  1. [1]

    Street tryon: Learning in-the-wild virtual try-on from unpaired person images,

    Aiyu Cui, Jay Mahajan, Viraj Shah, Preeti Gomathinayagam, Chang Liu, and Svetlana Lazebnik, “Street tryon: Learning in-the-wild virtual try-on from unpaired person images,” in CVPR, 2024

    work page 2024

  2. [2]

    Catvton: Concatenation is all you need for virtual try-on with diffusion models,

    Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, and Xiaodan Liang, “Catvton: Concatenation is all you need for virtual try-on with diffusion models,” arXiv preprint arXiv:2407.15886 , 2024

    arXiv 2024

  3. [3]

    M&m vto: Multi-garment virtual try-on and editing,

    Luyang Zhu, Yingwei Li, Nan Liu, Hao Peng, Dawei Yang, and Ira Kemelmacher-Shlizerman, “M&m vto: Multi-garment virtual try-on and editing,” in CVPR, 2024

    work page 2024

  4. [4]

    Mmtryon: Multi-modal multi- reference control for high-quality fashion generation,

    Xujie Zhang, Ente Lin, Xiu Li, Yuxuan Luo, Michael Kampffmeyer, Xin Dong, and Xiaodan Liang, “Mmtryon: Multi-modal multi- reference control for high-quality fashion generation,” arXiv preprint arXiv:2405.00448, 2024

    arXiv 2024

  5. [5]

    Tunnel try- on: Excavating spatial-temporal tunnels for high-quality virtual try-on in videos,

    Zhengze Xu, Mengting Chen, Zhao Wang, Linyu Xing, Zhonghua Zhai, Nong Sang, Jinsong Lan, Shuai Xiao, and Changxin Gao, “Tunnel try- on: Excavating spatial-temporal tunnels for high-quality virtual try-on in videos,” in ACM MM, 2024

    work page 2024

  6. [6]

    Vivid: Video virtual try-on using diffusion models,

    Zixun Fang, Wei Zhai, Aimin Su, Hongliang Song, Kai Zhu, Mao Wang, Yu Chen, Zhiheng Liu, Yang Cao, and Zheng-Jun Zha, “Vivid: Video virtual try-on using diffusion models,” arXiv preprint arXiv:2405.11794, 2024

    arXiv 2024

  7. [7]

    Generative adversarial networks,

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio, “Generative adversarial networks,” Communications of the ACM , 2020

    work page 2020

  8. [8]

    High-resolution image synthesis with latent diffu- sion models,

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer, “High-resolution image synthesis with latent diffu- sion models,” in CVPR, 2022

    work page 2022

Show all 42 references
  1. [9]

    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, 2018

    2018

  2. [10]

    Towards photo-realistic virtual try-on by adaptively generating-preserving image content,

    Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo, “Towards photo-realistic virtual try-on by adaptively generating-preserving image content,” in CVPR, 2020

    2020

  3. [11]

    High-resolution virtual try-on with misalignment and occlusion-handled conditions,

    Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo, “High-resolution virtual try-on with misalignment and occlusion-handled conditions,” in ECCV, 2022

    2022

  4. [12]

    Gp-vton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning,

    Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang, “Gp-vton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning,” in CVPR, 2023

    2023

  5. [13]

    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, 2024

    2024

  6. [14]

    Dragondiffusion: Enabling drag-style manipulation on diffusion mod- els,

    Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang, “Dragondiffusion: Enabling drag-style manipulation on diffusion mod- els,” in ICLR, 2024

    2024

  7. [15]

    Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing,

    Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang, “Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing,” in CVPR, 2024

    2024

  8. [16]

    Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on,

    Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo, “Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on,” in CVPR, 2024

    2024

  9. [17]

    Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on,

    Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen, “Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on,” arXiv preprint arXiv:2403.01779 , 2024

    2024 arXiv

  10. [18]

    Improving diffusion models for authentic virtual try-on in the wild,

    Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin, “Improving diffusion models for authentic virtual try-on in the wild,” in ECCV, 2025

    2025

  11. [19]

    Training language models to follow instructions with human feedback,

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wain- wright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al., “Training language models to follow instructions with human feedback,” in NeurIPS, 2022

    2022

  12. [20]

    Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al., “Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,” https://vicuna.lmsys.org, 2023

    2023

  13. [21]

    Qwen2 technical report,

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al., “Qwen2 technical report,” arXiv preprint arXiv:2407.10671 , 2024

    2024 arXiv

  14. [22]

    Llama: Open and efficient foundation language models,

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie- Anne Lachaux, Timoth´ee Lacroix, Baptiste Rozi `ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al., “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971 , 2023

    2023 arXiv

  15. [23]

    Editguard: Versatile image watermarking for tamper localiza- tion and copyright protection,

    Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang, “Editguard: Versatile image watermarking for tamper localiza- tion and copyright protection,” in CVPR, 2024

    2024

  16. [24]

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

    Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan, “Visual chatgpt: Talking, drawing and editing with visual foundation models,” arXiv preprint arXiv:2303.04671 , 2023

    2023 arXiv

  17. [25]

    Gpt4tools: Teaching large language model to use tools via self-instruction,

    Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan, “Gpt4tools: Teaching large language model to use tools via self-instruction,” in NeurIPS, 2024

    2024

  18. [26]

    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,” in NeurIPS, 2024

    2024

  19. [27]

    Gorilla: Large language model connected with massive apis,

    Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez, “Gorilla: Large language model connected with massive apis,” arXiv preprint arXiv:2305.15334, 2023

    2023 arXiv

  20. [28]

    React: Synergizing reasoning and acting in language models,

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao, “React: Synergizing reasoning and acting in language models,” in ICLR, 2023

    2023

  21. [29]

    Self-correction for human parsing,

    Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang, “Self-correction for human parsing,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020

    2020

  22. [30]

    Segment anything,

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al., “Segment anything,” in ICCV, 2023

    2023

  23. [31]

    Densepose: Dense human pose estimation in the wild,

    Rıza Alp G ¨uler, Natalia Neverova, and Iasonas Kokkinos, “Densepose: Dense human pose estimation in the wild,” in CVPR, 2018

    2018

  24. [32]

    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, 2023

    2023

  25. [33]

    Qwen2- vl: Enhancing vision-language model’s perception of the world at any resolution,

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al., “Qwen2- vl: Enhancing vision-language model’s perception of the world at any resolution,” arXiv preprint arXiv:2409.12191 , 2024

    2024 arXiv

  26. [34]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis,

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach, “Sdxl: Improving latent diffusion models for high-resolution image synthesis,” arXiv preprint arXiv:2307.01952 , 2023

    2023 arXiv

  27. [35]

    Viton-hd: High-resolution virtual try-on via misalignment-aware nor- malization,

    Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo, “Viton-hd: High-resolution virtual try-on via misalignment-aware nor- malization,” in CVPR, 2021

    2021

  28. [36]

    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, 2018

    2018

  29. [37]

    Learning transferable visual models from natural language supervision,

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al., “Learning transferable visual models from natural language supervision,” in ICML, 2021

    2021

  30. [38]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    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 equilibrium,” in NeurIPS, 2017

    2017

  31. [39]

    Instructpix2pix: Learning to follow image editing instructions,

    Tim Brooks, Aleksander Holynski, and Alexei A Efros, “Instructpix2pix: Learning to follow image editing instructions,” in CVPR, 2023

    2023

  32. [40]

    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

    2023

  33. [41]

    Ultraedit: Instruction-based fine-grained image editing at scale,

    Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang, “Ultraedit: Instruction-based fine-grained image editing at scale,” arXiv preprint arXiv:2407.05282, 2024

    2024 arXiv

  34. [42]

    Magic- quill: An intelligent interactive image editing system,

    Zichen Liu, Yue Yu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Wen Wang, Zhiheng Liu, Qifeng Chen, and Yujun Shen, “Magic- quill: An intelligent interactive image editing system,” arXiv preprint arXiv:2411.09703, 2024

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.