Pith. sign in

REVIEW 3 major objections 4 minor 78 references

Does Feasibility Matter? Understanding the Impact of Feasibility on Synthetic Training Data

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Feasibility of synthetic training images barely changes CLIP accuracy

desk verdict A useful controlled-generation pipeline and a plausible but not yet proven null result — feasibility barely matters for CLIP classification, but single runs and contaminated infeasible sets keep the claim conditional. read the letter →

arxiv 2505.10551 v1 pith:5LH52RYN submitted 2025-05-15 cs.CV cs.AI

Yiwen Liu , Jessica Bader , Jae Myung Kim This is my paper · ORCID
classification cs.CVcs.AI
keywords feasibilitysynthetictrainingdataCLIPLoRAdiffusionmodelsimageeditingfine-grainedclassificationaugmentation
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 asks whether generated training images must be 'feasible'—able to plausibly occur in the real world—to be useful for training a classifier. It answers no: when a CLIP classifier is fine-tuned with LoRA on synthetic images that differ from real photos only in one edited attribute, swapping feasible attributes for infeasible ones moves top-1 accuracy by less than 0.3 percent in most settings. The attribute edited matters more than feasibility: background changes improve accuracy whether feasible or infeasible, while color and texture changes help less and degrade beyond small scales. Mixed feasible/infeasible training sets perform about as well as pure sets. If the result holds, synthetic-data pipelines need not spend effort enforcing realism for every image.

What carries the argument

The VariReal pipeline generates minimal-change feasible/infeasible image pairs from real images. It uses GPT-4 with in-context examples to produce class-specific feasible and infeasible attribute prompts (background, color, texture), then performs prior-guided diffusion editing: inpainting with a real prior and mask dilation for backgrounds, and a two-stage inpainting-plus-ControlNet with Canny-edge conditioning for foreground color and texture edits. An MLLM (Llava-Next) filters outputs by attribute and feasibility. Performance is measured by LoRA fine-tuning of CLIP, computing $\Delta_1 = F - IF$ (feasible minus infeasible top-1 accuracy) and $\Delta_2 = \text{Mix} - \frac{F+IF}{2}$ (mixing gain over the average of pure sets). The mechanism that carries the argument is this pairing: because each real image is edited in only one attribute and every image is reused across all prompts, feasible and infeasible training sets differ only in the target attribute, isolating feasibility from other image content.

What would settle it

Train LoRA-fine-tuned CLIP on feasible and infeasible synthetic sets that have been censored by a human panel so every image is verified as correctly labeled, across many random seeds. If the top-1 accuracy gap between the two sets consistently exceeds 1% on any of the three attributes—or if the paper's below-0.3% differences fail to replicate under this stricter label control—then the claim that feasibility has no clear impact is falsified.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that feasibility has no clear impact on classification performance for LoRA-fine-tuned CLIP models. Across Oxford Pets, FGVC-Aircraft, and Stanford Cars, most differences $\Delta_1 = F - IF$ are below 0.3%, with the largest being a 1.8% gain for feasible backgrounds on Aircraft and a 3.2% gain for feasible textures on Cars; the direction is not consistent. The main corollary is that infeasible synthetic images can be used interchangeably with feasible ones: generation need not enforce feasibility, and mixing the two types does not hurt. The paper's second claim is that the attribute type matters: background edits consistently beat real-only training even when infeasible, whereas color and texture edits are less effective and may degrade performance at scale.

Load-bearing premise

The comparison assumes that 'feasible' and 'infeasible' are cleanly separated: the whole result rests on prompt generation, manual filtering, and per-attribute editing parameters producing truly infeasible images, yet human evaluation labeled up to 19.1% of infeasible texture images as feasible, so if infeasible data is actually in-distribution or the edits also change naturalness or class-relevant cues, the null finding becomes an artifact of the generator rather than a property of feasibility.

Editorial extensions

If this is right

  • Synthetic-data pipelines for object-centric classification can stop treating infeasible outputs as waste; both feasible and infeasible images are usable training material.
  • Background diversity is the more useful axis of augmentation: both feasible and infeasible background edits outperform real-only training, so constraints like ALIA's feasible-background restriction are unnecessary.
  • Color and texture edits should be used sparingly as augmentation; their benefit peaks at low synthetic-to-real ratios and turns negative at larger scale.
  • Mixing feasible and infeasible data is safe and can even add complementary signal in synthetic-only training, since the two types push models toward different correct-prediction sets while reaching similar accuracy.

Reading between the lines

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

  • Editorial inference: the paper's null result is probably a lower bound on feasibility effects: up to 19.1% of 'infeasible' texture images passed human inspection as feasible, so a cleaner operationalization of infeasibility could either shrink the gap further or reveal a genuine effect.
  • Editorial inference: the pattern that foreground edits underperform despite being distributionally closer to real data suggests classifiers rely on class-relevant foreground cues; feasibility research should separate 'distributional closeness' from 'task-relevance'.
  • Editorial inference: because the evidence comes from CLIP with LoRA and object-centric datasets, the conclusion may not transfer to other architectures, tasks like detection or segmentation, or safety-critical domains; testing with a convolutional baseline or a medical or autonomous-driving dataset would decide.
  • Editorial inference: the strongest testable extension is to vary one attribute at a time with a hard feasibility guarantee, such as physically impossible but photorealistic combinations verified by a panel, and measure per-seed accuracy gaps to see whether the below-0.3% null holds outside the paper's generator.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes VariReal, a diffusion-based minimal-edit pipeline that creates feasible and infeasible variants of real images along background, color, and texture attributes, and evaluates them on three fine-grained datasets (Oxford Pets, FGVC Aircraft, Stanford Cars) plus WaterBirds for background. CLIP models are LoRA-fine-tuned under synthetic-only and real+synthetic training, comparing feasible, infeasible, and mixed conditions. The main empirical claim is that feasibility has no clear impact on top-1 accuracy (Observation 1), that background edits help while color/texture edits are less effective (Observation 2), and that strictly feasible synthetic data is not necessary (Observation 3). The paper includes human evaluation, MLLM-based filtering, distributional metrics, scaling experiments, and ablations supporting the dataset-generation pipeline.

Significance. If the null result survives proper statistical treatment, the paper makes a useful and somewhat counterintuitive contribution: it challenges the common assumption that infeasible synthetic images are harmful and suggests that synthetic-data pipelines need not strictly enforce feasibility. The VariReal pipeline is a concrete, reusable tool for controlled attribute editing, with code released, and the study design includes human evaluation, automatic filtering, distributional analyses, and scaling/ablation experiments, which are strengths. However, the central evidence consists of small accuracy differences without variance estimates or significance testing, so the contribution is currently conditional on additional statistical validation.

major comments (3)
  1. [§4.2.1, Table 1, Observation 1] The central claim that feasibility 'shows no clear impact' rests on point estimates from a single LoRA fine-tuning run per condition; no seed variation, confidence intervals, or significance tests are reported. Most Δ1 values are below 0.3%, but per-cell gaps reach 3.2% (Cars, texture, synthetic-only), and the paper provides no estimate of run-to-run variability for this training protocol. Without repeated seeds or an equivalence test with a pre-specified margin, the absence of a visible gap cannot be distinguished from a real effect obscured by noise. Please add multi-seed results with variance reporting and a formal or semi-formal test of the no-effect claim.
  2. [§4.2.1] The paragraph reporting the distribution of Δ1 values contains an internal arithmetic error: after incorporating real data, the text states that '78% of Δ1 values (5/9) remain within 0.3%,' but 5/9 is 56%, not 78%. The sentence reports both 78% and the fraction 5/9, so the quantitative summary is internally inconsistent. This needs correction and a re-audit of all percentages and counts in the section.
  3. [§3.2.3, Table 2, Supp. G, Supp. Table 5] The operationalization of feasibility is not clean enough to support a strong null conclusion. Table 2 reports 80.9% feasibility correctness for infeasible texture images, so 19.1% of that training set is judged feasible by human evaluators; Supp. G attributes this to textures that only convey color and to implausible object-background combinations. If a substantial share of 'infeasible' data is actually in-distribution, any true feasibility effect is attenuated toward zero. Additionally, Supp. Table 5 shows that feasible and infeasible conditions use different generation hyperparameters (guidance scales, strengths, mask dilation/alpha factors), so the comparison varies in more than the target attribute. I request a contamination-filtered analysis on the human-verified subsets and either matched generation hyperparameters or a sensitivity analysis showing robustness to these choices.
minor comments (4)
  1. [Figure 4 caption] The caption contains a typo: 'Standford-Cars' should be 'Stanford Cars'.
  2. [Supp. Figure 13 caption] The caption begins with 'TQualitative results', which appears to be missing a leading word or contains a typo; please correct it.
  3. [§3.1] The notation ISyn is introduced in the task formulation but not defined at first use; please define it explicitly.
  4. [Table 1 caption] The formula for Δ2 is written as 'Mix− F+IF/2', which is ambiguous; it should be written as Mix − (F + IF)/2.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the feasibility comparison is an external empirical measurement, with only a minor non-load-bearing self-citation to DataDream for hyperparameter tuning.

full rationale

The paper contains no derivation chain whose conclusion is equivalent to its inputs. The central claim (Observation 1: 'feasibility shows no clear impact on classification performance') is an empirical measurement: top-1 accuracy is computed on the original held-out test sets of Oxford Pets, FGVC Aircraft, and Stanford Cars after LoRA fine-tuning, and the feasible/infeasible comparison is evaluated externally to the label-generation process. The feasible and infeasible prompt lists are curated (GPT-4 plus manual verification, Sec. 3.2.1) and further validated by human annotators (Table 2); this is construct validation, not a fitted parameter renamed as a prediction, because the reported accuracy numbers are not derived from the feasibility labels. The only self-citation with overlapping authors that carries any procedural weight is 'Following DataDream [28], we tune learning rates and weight decay' (Supp. B); this is a hyperparameter-tuning convention and is not load-bearing for the feasibility conclusion. The paper's own stated limitations—per-condition generation hyperparameters (Supp. A) and the 19.1% of infeasible texture images judged feasible by human evaluators (Table 2)—weaken the strength of the null result as a measurement claim, but they do not make the result circular. The reported arithmetic inconsistency in Sec. 4.2.1 (5/9 is 56%, not 78%) is an error in presentation, not a circularity. No equation in the paper reduces the target claim to its inputs, and no invoked theorem or prior result by the same authors is used to force the conclusion. Hence a score of 1 reflects only the minor, non-load-bearing self-citation.

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

This is an empirical study rather than a derivation. The central claim depends on two sets of hand-chosen degrees of freedom: per-condition generation parameters and manual/LLM prompt curation. They are not fitted to the classification outcome in the sense of a model, but they do shape the feasible-infeasible contrast and are not controlled for error.

free parameters (2)
  • Per-condition generation hyperparameters (guidance scales, strengths, mask dilation/alpha factors) = Varies per dataset and feasibility (Supplementary Table 5)
    Hand-chosen separately for each dataset and for feasible/infeasible conditions; differences in image quality between conditions could be confounded with feasibility.
  • CLIP fine-tuning hyperparameters (learning rate, weight decay) = lr from {1e-3, 5e-4, 1e-4, 5e-5, 1e-5}; weight decay from {1e-3, 1e-4, 5e-5}
    Selected per dataset on validation; without multiple seeds, small top-1 differences cannot be separated from tuning noise.
assumptions (2)
  • domain assumption Feasibility is a well-defined binary property for each attribute-class pair, and infeasible images are truly out-of-distribution for the real test set.
    Sec. 3.2.1 defines feasibility via GPT-4 and manual checks; Table 2 shows infeasible images are judged feasible up to 19.1% of the time, so the operationalization is noisy.
  • domain assumption Synthetic edits preserve all class-identifying and task-relevant information except the manipulated attribute.
    Sec. 3.2.2 copies invariant regions back, but Supp. Figure 19 shows some color edits lower naturalness; if class-relevant cues change, the attribute comparison is confounded.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Does Feasibility Matter? Understanding the Impact of Feasibility on Synthetic Training Data." pith.science (2026). https://pith.science/paper/5LH52RYN

@misc{pith2026250510551,
  author       = {Pith},
  title        = {Pith review of: Does Feasibility Matter? Understanding the Impact of Feasibility on Synthetic Training Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5LH52RYN}},
  note         = {Machine review of arXiv:2505.10551}
}
read the original abstract

With the development of photorealistic diffusion models, models trained in part or fully on synthetic data achieve progressively better results. However, diffusion models still routinely generate images that would not exist in reality, such as a dog floating above the ground or with unrealistic texture artifacts. We define the concept of feasibility as whether attributes in a synthetic image could realistically exist in the real-world domain; synthetic images containing attributes that violate this criterion are considered infeasible. Intuitively, infeasible images are typically considered out-of-distribution; thus, training on such images is expected to hinder a model's ability to generalize to real-world data, and they should therefore be excluded from the training set whenever possible. However, does feasibility really matter? In this paper, we investigate whether enforcing feasibility is necessary when generating synthetic training data for CLIP-based classifiers, focusing on three target attributes: background, color, and texture. We introduce VariReal, a pipeline that minimally edits a given source image to include feasible or infeasible attributes given by the textual prompt generated by a large language model. Our experiments show that feasibility minimally affects LoRA-fine-tuned CLIP performance, with mostly less than 0.3% difference in top-1 accuracy across three fine-grained datasets. Also, the attribute matters on whether the feasible/infeasible images adversarially influence the classification performance. Finally, mixing feasible and infeasible images in training datasets does not significantly impact performance compared to using purely feasible or infeasible datasets.

Figures

Figures reproduced from arXiv: 2505.10551 by the authors.

Figure 1
Figure 1. We propose VariReal, a pipeline for minimal-change editing of real images, enabling isolation of target attributes in three [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. We compare images generated by various candidate methods: Inpainting model [ [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An overview of VariReal pipeline. Minimal-change steps for background, color, and texture are highlighted in green, pink, and grey, respectively. Real images are processed to generate guidance maps (e.g., masks, Canny edges) for Inpainting and ControlNet. GPT-4 generates feasible and infeasible prompts (Pf and Pif ), which guide color retrieval or prior image generation via Stable Diffusion. These Real Priors, combi… view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Selected generation results from the three datasets. Only [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The FID score settings compared using feasible and in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: The generated attributes(prompt words) and self-filtering [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: A specific prompt example used for background prompt words generation of Oxford Pets [ [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Final accepted prompt word examples for Oxford Pets [ [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: The automatic filtering process using a MLLM model to filter the generated images using pre-defined qustions to check certain [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Qualitative results of the class Abyssinian from Oxford Pets dataset [ [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 13
Figure 13. Figure 13: TQualitative results of the class BMW X3 SUV 2012 [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 15
Figure 15. Figure 15: The averaged Inclusion and Jaccard index matrix for three editing settings across three datasets. ”f” = feasible, ”if” = infeasible, [PITH_FULL_IMAGE:figures/full_fig_p020_15.png]
Figure 16
Figure 16. Figure 16: Randomly selected generated samples across three datasets and feasibility attributes are shown. For visualization purposes, all [PITH_FULL_IMAGE:figures/full_fig_p021_16.png]
Figure 17
Figure 17. Figure 17: Instructions for feasibility and naturalness generated images human study. [PITH_FULL_IMAGE:figures/full_fig_p022_17.png]
Figure 18
Figure 18. Figure 18: Examples assessed as incorrect feasibility by human evaluators, including unclear fine-grained textures (e.g., ”fish scale”) and [PITH_FULL_IMAGE:figures/full_fig_p023_18.png]
Figure 19
Figure 19. Figure 19: Examples assessed by human evaluators as having lower naturalness, often due to unnatural color modifications or unrealistic [PITH_FULL_IMAGE:figures/full_fig_p023_19.png]
Figure 20
Figure 20. Figure 20: The ablation study for the usage to expand object mask for background edition setting. We show the real generated prior [PITH_FULL_IMAGE:figures/full_fig_p024_20.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

78 extracted references · 38 canonical work pages

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 ,

    arXiv

  2. [2]

    VisMin: Visual minimal-change understanding

    Rabiul Awal, Saba Ahmadi, Le Zhang, and Aishwarya Agrawal. VisMin: Visual minimal-change understanding. arXiv preprint arXiv:2407.16772, 2024. 3

    arXiv 2024

  3. [3]

    ClearDepth: en- hanced stereo perception of transparent objects for robotic manipulation

    Kaixin Bai, Huajian Zeng, Lei Zhang, Yiwen Liu, Hongli Xu, Zhaopeng Chen, and Jianwei Zhang. ClearDepth: en- hanced stereo perception of transparent objects for robotic manipulation. arXiv preprint arXiv:2409.08926, 2024. 3

    arXiv 2024

  4. [4]

    Deep learners benefit more from out-of-distribution examples

    Yoshua Bengio, Fr ´ed´eric Bastien, Arnaud Bergeron, Nicolas Boulanger-Lewandowski, Thomas Breuel, Youssouf Chher- awala, Moustapha Cisse, Myriam C ˆot´e, Dumitru Erhan, Jeremy Eustache, et al. Deep learners benefit more from out-of-distribution examples. In Proceedings of the Four- teenth International Conference on Artificial Intelligence and Statistics...

    work page 2011

  5. [5]

    Briaai background removal v1.4 model, 2024

    Briaai. Briaai background removal v1.4 model, 2024. https://huggingface.co/briaai/RMBG-1.4. 5

    work page 2024

  6. [6]

    In- structpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. In- structpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392–18402, 2023. 1, 3

    work page 2023

  7. [7]

    Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1

    2024

  8. [8]

    Diversified in-domain synthesis with efficient fine-tuning for few-shot classification

    Victor G Turrisi da Costa, Nicola Dall’Asen, Yiming Wang, Nicu Sebe, and Elisa Ricci. Diversified in-domain synthesis with efficient fine-tuning for few-shot classification. arXiv preprint arXiv:2312.03046, 2023. 2, 3

    arXiv 2023

Show all 78 references
  1. [9]

    The value of out-of- distribution data

    Ashwin De Silva, Rahul Ramesh, Carey Priebe, Pratik Chaudhari, and Joshua T V ogelstein. The value of out-of- distribution data. In International Conference on Machine Learning, pages 7366–7389. PMLR, 2023. 2

    2023

  2. [10]

    ImageNet: A large-scale hierarchical im- age database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical im- age database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 5

    2009

  3. [11]

    The MNIST database of handwritten digit images for machine learning research

    Li Deng. The MNIST database of handwritten digit images for machine learning research. IEEE signal processing mag- azine, 29(6):141–142, 2012. 2

    2012

  4. [12]

    A survey on in-context learning

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 4

    2022 arXiv

  5. [13]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint a...

    2010 arXiv

  6. [14]

    Gonzalez, and Trevor Darrell

    Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E. Gonzalez, and Trevor Darrell. Diversify your vi- sion datasets with automatic diffusion-based augmentation. NeurIPS, 36:79024–79034, 2023. 2, 3, 5, 6, 7, 1, 4

    2023

  7. [15]

    Deep genera- tive models for synthetic data: A survey

    Peter Eigenschink, Thomas Reutterer, Stefan Vamosi, Ralf Vamosi, Chang Sun, and Klaudius Kalcher. Deep genera- tive models for synthetic data: A survey. IEEE Access, 11: 47304–47320, 2023. 2

    2023

  8. [16]

    Scaling laws of synthetic images for model training

    Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382–7392, 2024. 3, 5

    2024

  9. [17]

    Instagen: Enhancing object detection by training on synthetic dataset

    Chengjian Feng, Yujie Zhong, Zequn Jie, Weidi Xie, and Lin Ma. Instagen: Enhancing object detection by training on synthetic dataset. In CVPR, pages 14121–14130, 2024. 3

    2024

  10. [18]

    Guiding instruction-based im- age editing via multimodal large language models

    Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based im- age editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 3

    2023 arXiv

  11. [19]

    How much data are augmentations worth? an in- vestigation into scaling laws, invariance, and implicit regu- larization

    Jonas Geiping, Micah Goldblum, Gowthami Somepalli, Ravid Shwartz-Ziv, Tom Goldstein, and Andrew Gordon Wilson. How much data are augmentations worth? an in- vestigation into scaling laws, invariance, and implicit regu- larization. arXiv preprint arXiv:2210.06441, 2022. 2

    arXiv

  12. [20]

    Synthetic data in health care: A narrative review

    Aldren Gonzales, Guruprabha Guruswamy, and Scott R Smith. Synthetic data in health care: A narrative review. PLOS Digital Health, 2(1):e0000082, 2023. 2

    2023

  13. [21]

    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. Commu- nications of the ACM, 63(11):139–144, 2020. 1

    2020

  14. [22]

    SynthCLIP: Are we Ready for a fully synthetic CLIP training? arXiv preprint arXiv:2402.01832, 2024

    Hasan Abed Al Kader Hammoud, Hani Itani, Fabio Pizzati, Philip Torr, Adel Bibi, and Bernard Ghanem. SynthCLIP: Are we Ready for a fully synthetic CLIP training? arXiv preprint arXiv:2402.01832, 2024. 2

    2024 arXiv

  15. [23]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016. 2

    2016

  16. [24]

    Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022

    Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 2, 3

    2022 arXiv

  17. [25]

    Prompt-to-prompt im- age editing with cross attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt im- age editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3

    2022 arXiv

  18. [26]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. NeurIPS, 33:6840–6851, 2020. 1, 2

    2020

  19. [27]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 5 9

    2021 arXiv

  20. [28]

    DataDream: Few-shot guided dataset generation

    Jae Myung Kim, Jessica Bader, Stephan Alaniz, Cordelia Schmid, and Zeynep Akata. DataDream: Few-shot guided dataset generation. arXiv preprint arXiv:2407.10910, 2024. 2, 3, 5, 1

    2024 arXiv

  21. [29]

    Diffblender: Scalable and composable multimodal text-to-image diffusion models

    Sungnyun Kim, Junsoo Lee, Kibeom Hong, Daesik Kim, and Namhyuk Ahn. Diffblender: Scalable and composable multimodal text-to-image diffusion models. arXiv preprint arXiv:2305.15194, 2023. 3

    2023 arXiv

  22. [30]

    Auto-encoding variational bayes

    Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 4

    2013 arXiv

  23. [31]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. In ICCV, pages 4015–4026, 2023. 1

    2023

  24. [32]

    3d object representations for fine-grained categorization

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV, pages 554–561, 2013. 5, 6, 4, 8

    2013

  25. [33]

    Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models

    Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 5

    2024 arXiv

  26. [34]

    Towards understanding cross and self-attention in stable diffusion for text-guided image editing

    Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In CVPR, pages 7817–7826, 2024. 1, 3

    2024

  27. [35]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 5

    2024

  28. [36]

    GroundingDino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. GroundingDino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5

    2023 arXiv

  29. [37]

    Summary of chatgpt-related research and perspective towards the future of large language models

    Yiheng Liu, Tianle Han, Siyuan Ma, Jiayue Zhang, Yuanyuan Yang, Jiaming Tian, Hao He, Antong Li, Meng- shen He, Zhengliang Liu, et al. Summary of chatgpt-related research and perspective towards the future of large language models. Meta-Radiology, page 100017, 2023. 1

    2023

  30. [38]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6, 1

    2017 arXiv

  31. [39]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461–11471, 2022. 3, 1, 2

    2022

  32. [40]

    Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and L Repaint Van Gool. Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461–11471, 2022. 3, 4

    2022

  33. [41]

    Fine-grained visual classi- fication of aircraft

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classi- fication of aircraft. arXiv preprint arXiv:1306.5151 , 2013. 5, 6, 7, 8, 4

    2013 arXiv

  34. [42]

    Backpropa- gating through fr\’echet inception distance

    Alexander Mathiasen and Frederik Hvilshøj. Backpropa- gating through fr\’echet inception distance. arXiv preprint arXiv:2009.14075, 2020. 6, 7

    2009 arXiv

  35. [43]

    Context diffusion: In-context aware image generation

    Ivona Najdenkoska, Animesh Sinha, Abhimanyu Dubey, Dhruv Mahajan, Vignesh Ramanathan, and Filip Radenovic. Context diffusion: In-context aware image generation. arXiv preprint arXiv:2312.03584, 5, 2023. 3

    2023 arXiv

  36. [44]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021. 3

    2021 arXiv

  37. [45]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6, 7

    2023 arXiv

  38. [46]

    Cats and dogs

    Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498–3505. IEEE, 2012. 5, 6, 3, 4, 7

    2012

  39. [47]

    Learning transferable visual models from natural language supervi- sion

    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 supervi- sion. In International conference on machine learning, ...

    2021

  40. [48]

    In- finigen indoors: Photorealistic indoor scenes using procedu- ral generation

    Alexander Raistrick, Lingjie Mei, Karhan Kayan, David Yan, Yiming Zuo, Beining Han, Hongyu Wen, Meenal Parakh, Stamatis Alexandropoulos, Lahav Lipson, et al. In- finigen indoors: Photorealistic indoor scenes using procedu- ral generation. In CVPR, pages 21783–21794, 2024. 3

    2024

  41. [49]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International confer- ence on machine learning, pages 8821–8831. Pmlr, 2021. 3

    2021

  42. [50]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5

    2024 arXiv

  43. [51]

    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. 1, 2, 3, 4, 5

    2022

  44. [52]

    U- Net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- Net: Convolutional networks for biomedical image segmen- tation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, pa...

    2015

  45. [53]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479–36494, 2022. 3

    2022

  46. [54]

    Fake it till you make it: Learning trans- ferable representations from synthetic imagenet clones

    Mert B ¨ulent Sarıyıldız, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning trans- ferable representations from synthetic imagenet clones. In CVPR, pages 8011–8021, 2023. 2, 3

    2023

  47. [55]

    Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using dif- fusion models

    Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using dif- fusion models. In ICCV, pages 1206–1217, 2023. 2, 3 10

    2023

  48. [56]

    Paint by example: Exemplar-based image editing with diffusion mod- els

    Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion mod- els. In CVPR, pages 18381–18391, 2023. 3

    2023

  49. [57]

    Generalized out-of-distribution detection: A survey

    Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. Inter- national Journal of Computer Vision, pages 1–28, 2024. 2

    2024

  50. [58]

    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

  51. [59]

    Diversify, don’t fine-tune: Scaling up visual recognition training with synthetic images

    Zhuoran Yu, Chenchen Zhu, Sean Culatana, Raghuraman Krishnamoorthi, Fanyi Xiao, and Yong Jae Lee. Diversify, don’t fine-tune: Scaling up visual recognition training with synthetic images. arXiv preprint arXiv:2312.02253, 2023. 2

    2023 arXiv

  52. [60]

    Dino: Detr with improved denoising anchor boxes for end-to-end object detection

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 1

    2022 arXiv

  53. [61]

    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, 4, 5, 1, 2

    2023

  54. [62]

    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,

  55. [63]

    Hive: Harnessing human feedback for instructional visual editing

    Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. In CVPR, pages 9026–9036,

  56. [64]

    a [CLS] in the [ATTRIBUTE] background

    Fan Zhou, Zengzhi Wang, Qian Liu, Junlong Li, and Pengfei Liu. Programming every example: Lifting pre- training data quality like experts at scale. arXiv preprint arXiv:2409.17115, 2024. 1 11 Does Feasibility Matter? Understanding the Impact of Feasibility on Synthetic Trainin...

    2024 arXiv

  57. [65]

    Unique [Attribute]: Ensure each listed [Attribute] is distinct and not synonymous with others pro- vided

  58. [66]

    Empty List Handling: If no unfeasible back- grounds can be identified, use ’EMPTY’ to denote this

  59. [67]

    Positive Example: • Object Class: [CLASS] • Question: Provide five different [Attribute] for the object class, each accompanied by a concise visual description

    Format Requirement: Answers must be formatted as a Python list, following the structure shown in the ’Answer’ section of the ’Example’. Positive Example: • Object Class: [CLASS] • Question: Provide five different [Attribute] for the object class, each accompanied by a concise ...

    2012

  60. [68]

    Feasible Backgrounds: Identify environments where the object class naturally occurs in the real world

  61. [69]

    inside a sun

    Unfeasible Backgrounds: Identify environments where the object class cannot naturally or logically be present. Avoid fantastical or scientifically impossible scenarios (e.g., "inside a sun")

  62. [70]

    Unique Backgrounds: Ensure each background is distinct and does not overlap in meaning with others

  63. [71]

    Empty List Handling: If no unfeasible backgrounds can be identified, use ’EMPTY’ to denote this

  64. [72]

    In Context Learning Example

    Format Requirement: Responses must be formatted as a Python list, following the structure provided in the ’Example’ section. In Context Learning Example

  65. [73]

    Question: Provide five different unfeasible backgrounds for a dog, each accompanied by a concise visual description

  66. [74]

    Answer: (a) ’underwater coral reef: A vibrant underwater scene filled with colorful corals, schools of fish, and shimmering light filtering through the water surface.’ (b) ’volcano crater: A rugged, rocky landscape with molten lava, steam vents, and an eerie red glow from the ...

  67. [75]

    The following answers are not acceptable: (a) ’industrial furnace room: A high-temperature environment with large furnaces used for metal smelting, filled with intense heat and noise.’ (b) ’operating theater: A sterile room in a hospital where surgeries are performed, requirin...

  68. [76]

    unsuitable for pets

    Reasons: (a) Responses are not in a proper Python list format (e.g., [’’, ’’, ..., ’’]). (b) Descriptions should focus on specific visual elements (e.g., objects, colors, lighting) instead of abstract concepts like "unsuitable for pets." (c) Example descriptions should include...

  69. [77]

    Please generate 20 different feasible and unfeasible backgrounds, respectively, for the class ’pets.’

  70. [78]

    edit instruction

    Additionally, provide detailed visual descriptions for each background. Figure 8. A specific prompt example used for background prompt words generation of Oxford Pets [46] dataset. 5 Feasible Prompt Word Examples from Pets Background: • suburban backyard: A grassy area with a ...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.