Pith. sign in

REVIEW 5 major objections 4 minor 35 references

A three-phase generative pipeline turns noisy factory photos into CAD-ready remnant contours.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A GAN-plus-VLM pipeline improves industrial remnant contour extraction, with GPT-image-1 outperforming Gemini 2.0 Flash on SSIM, LPIPS, and Hausdorff distance.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The paper's own Tables 1 and 2 show the VLM refinement stage degrades every metric, so the central three-phase claim is unsupported. the 5 major comments →

arxiv 2509.00284 v1 pith:QWUCNW55 submitted 2025-08-29 cs.CV cs.AI

Generative AI for Industrial Contour Detection: A Language-Guided Vision System

Liang Gong , Tommy (Zelin) Wang , Sara Chaker , Yanchen Dong , Fouad Bousetouane , Brenden Morton , Mark Mendez This is my paper
classification cs.CV cs.AI
keywords industrial computer visioncontour detectionremnant digitizationconditional GANvision-language model refinementprompt-guided image generationCAD-ready contoursmanufacturing automation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that remnant contour extraction in manufacturing—turning a noisy photo of a leftover sheet of metal, wood, or plastic into a usable outline—can be automated end-to-end by a three-phase generative pipeline: preprocessing, a fine-tuned Pix2Pix conditional GAN that produces rough binary contour masks, and a vision-language-model refinement stage in which a standardized text prompt guides image generation to clean and realign the contours. The authors argue that this beats both classical edge detectors and general-purpose foundation models on real factory imagery, and that the VLM refinement step is what closes the gap between plausible contours and CAD-level precision. If true, manufacturers could reuse off-cuts without slow manual tracing, and natural-language prompting would become a practical control channel for geometric editing in industrial vision.

Core claim

The paper's central claim: a three-phase pipeline—data preprocessing, a fine-tuned Pix2Pix GAN producing binary contour masks, and a multimodal stage that feeds the mask plus a standardized text prompt to GPT-image-1—yields CAD-level remnant contours. Pix2Pix alone scored high on perceptual metrics (SSIM 0.9705, LPIPS 0.0530) but drifted geometrically (Hausdorff 58.22, IoU 0.6265). With identical prompts and masks, GPT-image-1 beat Gemini 2.0 Flash on structural and perceptual quality (SSIM 0.9044 vs 0.6525; Hausdorff 98.046 vs 161.353). Conclusion: prompt-guided refinement in a human-in-the-loop workflow closes the gap between automated contour generation and CAD requirements.

What carries the argument

The carrying mechanism is the multimodal refinement prompt: a standardized natural-language instruction paired with the Pix2Pix output mask, passed as combined image-plus-text input to an image generation model (GPT-image-1) that edits rather than only captions. The prompt is crafted through a human-in-the-loop process (Section 3.5) and applied identically to both API backends, making the prompt the controlled variable in the head-to-head benchmark. This mechanism supplies the geometric regularization the GAN alone cannot: it fills missing contour gaps and realigns edges toward ground truth.

Load-bearing premise

The whole head-to-head benchmark rests on one unverifiable premise: the single standardized prompt, developed through a human-in-the-loop process, was neutral and did not favor GPT-image-1 over Gemini 2.0 Flash—and because the prompt is not disclosed, that neutrality cannot be checked from the paper.

What would settle it

Rerun both APIs on the same 274 test images with several prompt variants, including one written by a person blind to which models are compared, and check whether GPT-image-1's lead in SSIM, LPIPS, and Hausdorff distance persists; if the ranking shrinks or flips across prompts, the reported superiority is prompt-specific rather than intrinsic to the model.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Manual tracing becomes replaceable for remnant reuse: an operator photographs a leftover sheet and the pipeline returns a CAD-ready contour, with the chatbot handling only exceptional cases.
  • The benchmark implies that structural editing ability varies widely across API-accessible image generation models: GPT-image-1 preserved contour geometry while Gemini 2.0 Flash introduced distortions, so geometric capability cannot be assumed from multimodal reasoning strength.
  • Perceptual-similarity metrics (SSIM, LPIPS) are not enough to certify an industrial contour: Pix2Pix scored high on both while still drifting geometrically, so CAD readiness must be tracked with geometric measures like Hausdorff distance and IoU.
  • The two-stage division of labor—cheap generator plus prompt-guided refiner—means precision does not have to be learned by a single model, so the refinement backend can be upgraded independently of the GAN stage.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same standardized refinement prompt could be applied to masks from other detectors—SAM, DINO, or classical edge finders—which would test whether prompt-guided refinement transfers beyond Pix2Pix outputs; the paper does not run that experiment.
  • Operator corrections entered through the chatbot are a natural training set for the active-learning loop the paper lists as future work, because that correction log already exists in the deployed system.
  • Releasing the exact prompt text and its revision history would let outsiders rebuild the head-to-head benchmark with a prompt grid, separating what is intrinsic to GPT-image-1 from what is tuned into the prompt.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper proposes a three-phase generative vision system for industrial remnant contour detection. Phase 1 preprocesses and augments proprietary FabTrack images. Phase 2 fine-tunes a Pix2Pix conditional GAN to produce binary contour masks from noisy factory images. Phase 3 refines these masks with a vision-language model (primarily OpenAI's GPT-image-1) using standardized text prompts, optionally through a human-in-the-loop chatbot. The authors report quantitative results: fine-tuned Pix2Pix alone achieves SSIM 0.9705, LPIPS 0.0530, Hausdorff distance 58.22, and IoU 0.6265 (Table 1). On a 274-image test subset, GPT-image-1 refined outputs achieve SSIM 0.9044, LPIPS 0.2315, Hausdorff 98.046, IoU 0.6420, while Gemini 2.0 Flash achieves worse scores on all but IoU (Table 2). The paper concludes that the three-phase system significantly reduces manual tracing effort, improves accuracy, and produces contours with 'CAD-level precision.'

Significance. If the claims were supported, a language-guided generative pipeline that turns noisy factory-floor images into CAD-ready contours would be practically valuable for manufacturing digitization, nesting, and remnant reuse. The paper has some strengths: it evaluates two commercial VLM APIs on real proprietary factory data under a common prompt, and it reports quantitative image-similarity metrics alongside qualitative overlays. However, the core empirical claim—that VLM refinement improves geometric accuracy over Phase 2 alone—is contradicted by the paper's own numbers. The absence of a matched-pair comparison, the undisclosed standardized prompt, and the lack of any engineering tolerance metric leave the headline claims unsupported. As a result, the significance of the contribution cannot be assessed from the evidence presented.

major comments (5)
  1. [§4.2–4.3, Tables 1 and 2]
  2. [Abstract, §1.3, §3.4, Conclusion]
  3. [§3.4, §3.5, §4.3]
  4. [§4.3, Table 2]
  5. [§4.1, §1.2]
minor comments (4)
  1. [§2.1]
  2. [References [7] and [8]]
  3. [Throughout]
  4. [§4.4, Figures 4–5]

Circularity Check

0 steps flagged

No circular derivation: the central benchmark is direct and externally grounded; only minor non-load-bearing self-citation.

full rationale

The paper's central empirical claims are a direct comparison of two external commercial APIs (GPT-image-1 and Gemini 2.0 Flash) under a shared prompt, and a fine-tuned Pix2Pix evaluation against ground-truth contours on proprietary FabTrack data. These are not derived from the paper's own definitions or fitted parameters renamed as predictions. The only self-citation that structures the paper is the 'four-pillar taxonomy' of Bousetouane [4], a co-author's preprint, used in Section 2 to organize prior work: 'Recent advances in generative AI for vision can be systematically categorized using the four-pillar taxonomy introduced by Bousetouane [4].' This taxonomy is organizational background and future-work motivation ([3], [5]); it does not supply the system's components or the benchmark's ground truth, so it is not load-bearing. The undisclosed standardized prompt and the human-in-the-loop prompt development raise a fairness question for the GPT-image-1 vs Gemini comparison, but no specific reduction can be exhibited from the paper's text, so this is a correctness/transparency concern, not circularity. Similarly, the apparent metric regression between Table 1 and Table 2 (SSIM 0.9705 vs 0.9044, Hausdorff 58.22 vs 98.05) suggests the Phase 3 'improvement' claim is unsupported by the reported numbers, but that is an internal-consistency issue rather than a circular derivation. Accordingly, no significant circularity is found.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The central claim rests on the proprietary FabTrack dataset and its expert annotations, on the belief that perceptual and geometric similarity metrics reflect CAD-level precision, and on the neutrality of the undisclosed standardized prompt used for both VLM APIs. The paper introduces no new free numeric parameters beyond the hand-tuned (but undisclosed) prompt and hyperparameters; no invented entities are postulated. The evidence quality is constrained by unavailability of data, code, and protocol details.

free parameters (4)
  • Standardized prompt text = Not disclosed
    The prompt passed to both APIs is the key experimental variable; its content is hand-selected and undisclosed (Sec. 3.4, 4.3).
  • Pix2Pix hyperparameters = Not reported
    Learning rate, L1 weight, GAN mode, generator depth, normalization, batch size were tuned but values are omitted (Sec. 3.3).
  • Data augmentation magnitudes = Not reported
    Horizontal flips, random rotations, brightness shifts are mentioned without ranges (Sec. 3.2, 4.1).
  • Human-in-the-loop prompt adjustments = Not quantified
    The chatbot interface allows experts to iteratively change prompts (Sec. 3.5), but the number and nature of adjustments for the standardized prompt are not described.
axioms (3)
  • domain assumption The proprietary FabTrack dataset is representative of real-world manufacturing remnants and the expert annotations are accurate ground truth.
    All evaluation numbers in Tables 1 and 2 are computed against this unverifiable dataset (Sec. 4.1).
  • domain assumption The metrics SSIM, LPIPS, Hausdorff distance, and IoU are appropriate proxies for 'CAD-level precision' and 'geometric fidelity'.
    The paper equates these image-similarity metrics with CAD-level engineering precision (Sec. 4.2, 4.3), but no engineering tolerance is defined.
  • ad hoc to paper The undisclosed standardized prompt is equally fair to both VLM APIs and reflects the authors' human-in-the-loop process without bias.
    The exact prompt is never given, and it was iterated with human feedback (Sec. 3.4, 3.5), which could have been tuned for GPT-image-1.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative AI for Industrial Contour Detection: A Language-Guided Vision System." pith.science (2026). https://pith.science/paper/QWUCNW55

@misc{pith2026250900284,
  author       = {Pith},
  title        = {Pith review of: Generative AI for Industrial Contour Detection: A Language-Guided Vision System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QWUCNW55}},
  note         = {Machine review of arXiv:2509.00284}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Industrial computer vision systems often struggle with noise, material variability, and uncontrolled imaging conditions, limiting the effectiveness of classical edge detectors and handcrafted pipelines. In this work, we present a language-guided generative vision system for remnant contour detection in manufacturing, designed to achieve CAD-level precision. The system is organized into three stages: data acquisition and preprocessing, contour generation using a conditional GAN, and multimodal contour refinement through vision-language modeling, where standardized prompts are crafted in a human-in-the-loop process and applied through image-text guided synthesis. On proprietary FabTrack datasets, the proposed system improved contour fidelity, enhancing edge continuity and geometric alignment while reducing manual tracing. For the refinement stage, we benchmarked several vision-language models, including Google's Gemini 2.0 Flash, OpenAI's GPT-image-1 integrated within a VLM-guided workflow, and open-source baselines. Under standardized conditions, GPT-image-1 consistently outperformed Gemini 2.0 Flash in both structural accuracy and perceptual quality. These findings demonstrate the promise of VLM-guided generative workflows for advancing industrial computer vision beyond the limitations of classical pipelines.

Figures

Figures reproduced from arXiv: 2509.00284 by Brenden Morton, Fouad Bousetouane, Liang Gong, Mark Mendez, Sara Chaker, Tommy (Zelin) Wang, Yanchen Dong.

Figure 1
Figure 1. Figure 1: Examples of remnant images captured at the FabTrack Lab [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Four-pillar taxonomy of input-driven generative AI for vision, [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed three-phase generative system for in [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of contour refinement methods. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overlay visualization of contour alignment. Top: Ground truth [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 25 canonical work pages · 3 internal anchors

  1. [1]

    Moondream: Efficient open-source vision–language foundation model

    Moondream AI. Moondream: Efficient open-source vision–language foundation model. https://github.com/moondream-ai, 2024. Ac- cessed: 2025-08-25

    work page 2024

  2. [2]

    Qwen-vl: A frontier large-scale vision–language model with comprehensive capabilities

    Junjie Bai et al. Qwen-vl: A frontier large-scale vision–language model with comprehensive capabilities. arXiv preprint arXiv:2308.12966, 2023

    Pith/arXiv arXiv 2023

  3. [3]

    Agentic systems: A guide to transforming indus- tries with vertical ai agents

    Fouad Bousetouane. Agentic systems: A guide to transforming indus- tries with vertical ai agents. arXiv preprint arXiv:2501.00881, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Generative AI for Vision: A Comprehensive Study of Frameworks and Applications

    Fouad Bousetouane. Generative ai for vision: A comprehensive study of frameworks and applications. arXiv preprint arXiv:2501.18033, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    Information-theoretic Bayesian Optimization: Survey and Tutorial

    Fouad Bousetouane. Physical ai agents: Integrating cognitive intelli- gence with real-world action. arXiv preprint arXiv:2502.06789, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv´ e J´ egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650–9660, 2021

    work page 2021

  7. [8]

    Spherical Double K-Means: a co-clustering approach for text data analysis

    Yifan Chen, Zhuang Liu, Hao Wang, et al. Deepseek janus-pro: Unified multimodal foundation model for understanding and generation. arXiv preprint arXiv:2501.04562, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  8. [9]

    Bert: Pre-training of deep bidirectional transformers for language under- standing

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics, pages 4171– 4186, 2019. 17

    work page 2019

  9. [10]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780–8794, 2021

    work page 2021

  10. [11]

    Digital twin: Enabling technologies, challenges and open research

    Anthony Fuller, Zhidong Fan, Charles Day, and Chris Barlow. Digital twin: Enabling technologies, challenges and open research. IEEE Access, 8:108952–108971, 2020

    work page 2020

  11. [12]

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

    Tianxing Gao, Yuming Ma, Jian Zhang, et al. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023

    Pith/arXiv arXiv 2023

  12. [13]

    Gen- erative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Gen- erative adversarial nets. In Advances in Neural Information Processing Systems, pages 2672–2680, 2014

    work page 2014

  13. [14]

    Gemini 2.5: Deep Think Model Card, 2025

    Google DeepMind. Gemini 2.5: Deep Think Model Card, 2025. Model Card published August 1, 2025, detailing model capabilities, limitations, and evaluation

    work page 2025

  14. [15]

    beta-vae: Learning basic visual concepts with a constrained variational framework

    Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International Conference on Learning Representations, 2017

    2017

  15. [16]

    Denoising diffusion proba- bilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion proba- bilistic models. In Advances in Neural Information Processing Systems, pages 6840–6851, 2020

    work page 2020

  16. [17]

    Comparing images using the hausdorff distance

    Daniel P Huttenlocher, Gregory A Klanderman, and William J Ruck- lidge. Comparing images using the hausdorff distance. IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 15(9):850–863, 1993

    work page 1993

  17. [18]

    Image- to-image translation with conditional adversarial networks

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image- to-image translation with conditional adversarial networks. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 5967–5976, 2017. 18

    work page 2017

  18. [19]

    A style-based genera- tor architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based genera- tor architecture for generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4401–4410, 2019

    work page 2019

  19. [20]

    Auto-encoding variational bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

    Pith/arXiv arXiv 2013

  20. [21]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Camille Rol- land, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, Piotr Doll´ ar, and Ross Girshick. Segment anything. arXiv preprint arXiv:2304.02643, 2023

    Pith/arXiv arXiv 2023

  21. [22]

    Composable dif- fusion: Unlocking compositional creativity with text-to-image diffusion models

    Jiatao Liu, Xingqian Zhang, Fangzhou Hong, et al. Composable dif- fusion: Unlocking compositional creativity with text-to-image diffusion models. arXiv preprint arXiv:2302.03027, 2023

    Pith/arXiv arXiv 2023

  22. [23]

    Pose guided person im- age generation

    Liqian Ma, Qianru Sun, Zhen Li, and Jiaya Jia. Pose guided person im- age generation. In Advances in Neural Information Processing Systems, pages 406–416, 2017

    work page 2017

  23. [24]

    Gpt-image-1: Openai’s multimodal foundation model for image editing and reasoning

    OpenAI. Gpt-image-1: Openai’s multimodal foundation model for image editing and reasoning. https://openai.com/research, 2024. Accessed: 2025-08-25

    work page 2024

  24. [25]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth´ ee Darcet, Theo Moutakanni, Hugo B´ erard, Wei Li, Pierre Stock Mazare, Daniel Haziza, Thomas Hollebecque, Wojciech Galuba, Piotr Bojanowski, Mahmoud Assran, Ishan Misra, Armand Joulin, Herve Jegou, and Edouard Grave. Dinov2: Learning robust visual features without supervision. In Proceedings of the IEEE/CVF Conference on Computer Vi...

    work page 2023

  25. [26]

    Learning transfer- able 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, Gretchen Krueger, and Ilya Sutskever. Learning transfer- able visual models from natural language supervision. In International Conference on Machine Learning, pages 8748–8763, 2021

    work page 2021

  26. [27]

    Language models are unsupervised multitask learn- ers

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learn- ers. Technical report, OpenAI, 2019. 19

    work page 2019

  27. [28]

    Hierarchical text-conditional image generation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022

    Pith/arXiv arXiv 2022

  28. [29]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022

    work page 2022

  29. [30]

    Learning structured output representation using deep conditional generative models

    Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems, pages 3483–3491, 2015

    work page 2015

  30. [31]

    Neural discrete representation learning

    Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems, pages 6306–6315, 2017

    work page 2017

  31. [32]

    Image quality assessment: From error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600–612, 2004

    work page 2004

  32. [33]

    Holistically-nested edge detection

    Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1395–1403, 2015

    work page 2015

  33. [34]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023

    Pith/arXiv arXiv 2023

  34. [35]

    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 Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586–595, 2018

    2018

  35. [36]

    Unpaired image-to-image translation using cycle-consistent adversarial networks

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223–2232, 2017. 20

    work page 2017

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.