Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

Direct Ascent Synthesis: Revealing Hidden Generative Capabilities in Discriminative Models

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read Discriminative models like CLIP can generate natural images from text by gradient ascent alone, without any generative training.

desk verdict DAS is a neat optimization trick, but the central claim about hidden generative capability is under-supported without a random-encoder control. read the letter →

arxiv 2502.07753 v1 pith:MRH743WV submitted 2025-02-11 cs.CV

classification cs.CV
keywords CLIPmulti-resolutionoptimizationtraining-freeimagegenerationdiscriminativemodelsmodelinversionadversarialexamplesnaturalstatisticstext-to-image
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 attempts to establish that a pretrained discriminative model such as CLIP already contains enough generative knowledge to synthesize recognizable, natural-looking images from text descriptions or reference embeddings, with no generative training. The proposed method, Direct Ascent Synthesis (DAS), maximizes CLIP embedding similarity through gradient ascent over a multi-scale image parameterization rather than over raw pixels. The multi-scale form is what separates natural images from the noise-like adversarial patterns that ordinary inversion produces. If the claim holds, the distinction between discriminative and generative architectures is a matter of how representations are accessed, not what they contain. The paper supports the claim with text-to-image generation, embedding-based reconstruction, style transfer, and inpainting experiments, and with power spectra showing the outputs follow the $1/f^2$ statistics of natural images.

What carries the argument

The load-bearing object is the multi-resolution parameterization of Eq. (4): $$I = \tfrac{1}{2} + \tfrac{1}{2}\tanh\!\Bigl(\sum_{r\in\rho} \mathrm{resize}_{224}(P_r)\Bigr),$$ with components $P_r$ at resolutions from $1\times1$ to $224\times224$. Optimizing these components jointly, with random shifts and pixel noise as augmentations and an ensemble of CLIP models, distributes gradients across scales and suppresses the high-frequency solutions that characterize adversarial inversion. The parameterization acts as a prior that steers optimization toward the natural-image manifold; the paper shows the resulting power spectrum has slope close to $-2$, matching natural image statistics.

What would settle it

Run DAS with only a single high-resolution component (for example, $r=224$ only, so the sum collapses to direct pixel optimization) while keeping all other settings identical: the paper itself shows this yields noise-like adversarial patterns, confirming the prior carries the effect. A sharper test would apply DAS to a discriminative model trained on images with non-natural statistics; if natural-looking $1/f^2$ images still appear, the prior is not the mechanism claimed, and if they disappear, the central claim depends entirely on this parameterization.

Watch

Extended reading notes

Core claim

The central discovery is that inverting a CLIP embedding by direct gradient ascent becomes a viable generative procedure when the optimization variable is a sum of rescaled components at resolutions from $1\times1$ to $224\times224$, passed through a tanh nonlinearity to keep pixel values in range. Under this parameterization, the same objective that normally yields unrecognizable adversarial images instead produces semantically coherent scenes that track the prompt, preserve global composition, and follow the $1/f^2$ power spectrum of natural images. The paper interprets this as evidence that discriminative models encode generative structure as a byproduct of discriminative training, and that the historical failure of inversion was an optimization failure rather than an absence of generative information.

Load-bearing premise

The load-bearing premise is that the multi-resolution parameterization alone keeps gradient ascent within the natural-image manifold, preventing the adversarial noise solutions that direct pixel inversion produces; the paper concedes in Section 6.4 that it lacks a complete theoretical framework for why this works.

Editorial extensions

If this is right

  • Text-to-image generation can be performed with a frozen discriminative model and a few seconds of GPU time, bypassing the training cost of GANs and diffusion models.
  • Model inversion and adversarial example construction become two regimes of the same optimization, distinguished by whether the search respects natural image statistics.
  • Discriminative models that were never trained to generate can be reused for reconstruction, style transfer, inpainting, and structured patterns without fine-tuning.
  • The observed $1/f^2$ spectra of generated images suggest the method output lies on the natural-image manifold, opening a concrete link between generation quality and natural image statistics.
  • Ensembling CLIP models and using simple augmentations improves quality, so the generative capability scales with the richness of the discriminative representation rather than with generative architecture.

Reading between the lines

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

  • If the access story is right, the same multi-scale optimization should work on other embedding models trained on natural images, so a cheap test is to apply DAS to a non-CLIP discriminative vision model and measure whether generation quality tracks the model's semantic quality.
  • The dependence on a hand-chosen prior suggests a continuum: stronger scale-space priors may close the remaining quality gap to diffusion models without adding training, and the paper's concession that a theory is missing marks this as the place to look.
  • Because the paper reports that some CLIP models are unexpectedly bad generators for no obvious reason, the method could serve as an interpretability probe: model properties that resist DAS might reveal representations that did not learn natural image structure.
  • The implication that robust models may be inherently better generators is left open; a testable extension would be to correlate a model's adversarial robustness with DAS generation quality across checkpoints.
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

4 major / 5 minor

Summary. The paper proposes Direct Ascent Synthesis (DAS), a training-free method that generates images by optimizing a multi-resolution sum-of-components parameterization (Eq. 4) to maximize cosine similarity between a CLIP embedding of the resulting image and a target text or image embedding. The authors report qualitative results across text-to-image generation, style transfer, reconstruction, inpainting, and flag generation, and claim that the success of this simple optimization procedure demonstrates that discriminative models 'inherently contain powerful generative capabilities,' thereby challenging the traditional discriminative/generative dichotomy. The paper also claims that the generated images maintain natural-image statistics (a 1/f^2 power spectrum) and that the multi-resolution parameterization serves as a prior that prevents adversarial solutions. The central conceptual claim, however, is supported only by qualitative examples and one power-spectrum plot, with no quantitative evaluation, baselines, or controls that would separate the contribution of the parameterization from that of the model's representations.

Significance. If the central claim were established, the paper would be significant: it would show that standard discriminative models can serve as generators without any generative training, with implications for interpretability, adversarial robustness, and model reuse. The method itself is simple and the qualitative images are often compelling, suggesting a potentially useful optimization technique. The paper is also commendably explicit about its lack of a theoretical account (Section 6.4). However, the significance as stated depends on disentangling the effect of the multi-resolution parameterization from the properties of the discriminative model, and the current evidence does not achieve that separation. The claim that discriminative models 'inherently contain' generative capabilities requires a control that the manuscript does not provide; without it, the results are equally consistent with a hand-crafted spectral prior doing most of the work.

major comments (4)
  1. [Section 4.2, Eq. (4)] Eq. (4) is not a manifold constraint: because the resolution set rho includes r=224, setting all other P_r to zero lets P_224 alone parameterize every image in (0,1)^{224*224*3} via P_224 = atanh(2I-1). The paper's claim that the parameterization 'guides optimization to the natural image manifold' (Section 5.1) and 'suppresses high-frequency adversarial patterns' (Section 4.2) is therefore an empirical assertion, not a consequence of the parameterization. The only evidence offered is one power-spectrum plot (Fig. 2), which is insufficient because summing independent random components at multiple scales also produces approximately 1/f^2 spectra. The authors should provide a control experiment in which the same multi-resolution parameterization is optimized against a random encoder or a fixed random target; if such a control also yields natural-looking images, the central claim about discriminative models is unsupported.
  2. [Section 5 (all subsections)] The experimental evaluation is entirely qualitative. No quantitative metrics (e.g., CLIP score, FID/IS, or human-judgment statistics), no success rates, no per-seed or per-prompt variance measures, and no comparisons to baselines (e.g., direct pixel inversion with the same augmentations, or standard inversion methods such as Mahendran & Vedaldi 2014) are reported. Claims such as 'semantic consistency' and 'compositional understanding' in Sections 5.1 and 5.2 are based on selected examples shown in Figures 7 and 8. The paper needs a quantitative or at least a systematically sampled evaluation to support the strong statement in the abstract that these models 'inherently contain powerful generative capabilities.'
  3. [Section 4.2, Fig. 2] The reported 1/f^2 power spectrum is used as evidence that DAS produces natural image statistics, but no error bars, no confidence intervals, and no comparison to a baseline (e.g., a random-encoder control under the same parameterization) are given. Given that the paper concedes in Section 6.4 that it 'lack[s] a complete theoretical framework' for why multi-resolution optimization prevents adversarial solutions, the assertion in Section 4.2 that 'high-frequency adversarial patterns are suppressed by scale decomposition' is not established. The authors should either provide a formal argument or, more realistically, report the power spectrum for a control optimization that does not use a discriminative model.
  4. [Abstract and Section 6.1] The paper conflates the properties of the optimization parameterization with the properties of the discriminative model. The objective in Eq. (5) maximizes CLIP cosine similarity, so matching the embedding is built in; the real question is whether the resulting images are natural and semantic because of the model's representations or because of the multi-resolution prior. The claim that 'discriminative models inherently contain rich generative knowledge' (Section 6.1) requires a control that removes the model's learned representations (e.g., a randomly initialized or fixed random encoder) while keeping the parameterization, augmentations, and ensemble unchanged. Without such a control, the 'inherent' claim is circular. This is a load-bearing point for the paper's central thesis and must be addressed before the conceptual claim can be evaluated.
minor comments (5)
  1. [Section 2.3] Several citations are non-archival blog posts (Fort 2021a, 2021b, 2022); where peer-reviewed or arXiv versions exist, they should be cited instead, or the preprint status should be made explicit.
  2. [Figure 2] The power-spectrum plot lacks axis labels, units, and any indication of the fitting procedure used to obtain the reported slope of approximately -2; the authors should also state how many images were averaged and report error bars.
  3. [Section 4.3] The claim that the two augmentations (random x-y shifts and pixel noise) are 'crucial' and that 'neither is sufficient alone' is not supported by an ablation isolating their individual and joint contributions; the ablation in Figure 4 shows only three configurations and does not separate these two augmentations.
  4. [Section 5.5] The style-transfer comparison with raw-pixel optimization is described as 'somewhat subjective' and is not quantified; a simple perceptual metric or a small user study would strengthen the claim that DAS produces more pleasing results with fewer high-frequency artifacts.
  5. [Throughout] There are minor typographical errors (e.g., 'V AEs' in Section 2.1) and some claims are stated more strongly than the evidence supports; for example, Section 5.1 states the multi-resolution prior 'successfully constrains the optimization to the natural image manifold' without the controls discussed above.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DAS is an empirical optimization method whose matching objective is explicit, and self-citations are contextual rather than load-bearing.

full rationale

The paper's derivation chain is not circular. The optimization objective (Eq. 5) is CLIP cosine similarity (Eq. 3), and the multi-resolution parameterization (Eq. 4) is an explicit input representation; matching the target embedding is the stated goal, not a hidden premise. The claim that discriminative models 'inherently contain powerful generative capabilities' is an empirical interpretation of the synthesis results, and the naturalness of the outputs is assessed externally by human perception, not defined by the objective. The self-citations (Fort & Lakshminarayanan 2024; Fort 2025; Fort et al. 2022) appear in related work, motivation, and future directions; none is invoked as a uniqueness theorem or as the source of the method's validity, so they are not load-bearing. The paper explicitly concedes in Sec. 6.4 that 'we lack a complete theoretical framework' for why multi-resolution optimization prevents adversarial solutions; this is an honest limitation, not evidence of circularity. The strongest potential concern—that the reported 1/f^2 spectrum in Fig. 2 may be an artifact of the multi-resolution sum rather than evidence of natural-image guidance—is a methodological confound and an empirical controls question, not a case of the conclusion being equivalent to the input by construction. Because no fitted parameter is renamed as a prediction and no result is forced by self-citation, the circularity burden is not met.

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

The method relies on hand-selected optimization hyperparameters, corrective prompt weights, and an arbitrary resolution ladder; these are not derived and partially determine the visual quality. The conceptual claim rests on domain assumptions about CLIP embeddings and natural image statistics. No new physical entities or model components are postulated.

free parameters (4)
  • Corrective prompt weights (per application) = -0.6 to -0.3 for OCR, 0.3 for render quality, 0.6 for cohesive single subject
    Hand-tuned to suppress text and stacking artifacts in Figures 3, 7, and 10; no sensitivity analysis is reported, and these weights affect the perceived quality.
  • Optimization hyperparameters = 100 steps, SGD lr 2e-1, noise std 0.2, xy jitter +/-56, 32 augmentations
    Chosen by hand in Section 4.3 without ablations or sensitivity studies, so it is unclear which values the results depend on.
  • Resolution set rho = 1x1 through 224x224
    The set of scales is not derived from a principle; changing it could change the output power spectrum and image quality.
  • CLIP ensemble members = OpenAI ViT-B/32 plus two OpenCLIP ViT-B/32 variants, exact variants unspecified
    The ensemble 'marginally' improves quality, but some CLIP models are said to be 'particularly bad at being turned into generators' without explanation; this selection is ad hoc.
assumptions (4)
  • domain assumption CLIP embeddings encode sufficient semantic and stylistic information for image reconstruction and style transfer.
    Invoked in Sections 4.4 and 5.3; if false, DAS cannot recover content from embeddings.
  • ad hoc to paper The multi-resolution sum-of-rescaled-components parameterization is a natural-image prior that prevents adversarial patterns.
    Core mechanism asserted in Section 4.2 and Fig. 4; the paper states in Section 6.4 that no complete theoretical explanation exists.
  • ad hoc to paper Cosine similarity to the final-layer CLIP embedding is an adequate objective for perceptual quality.
    Defined in Eq. 3; no quantitative image-quality metrics validate this choice.
  • domain assumption Natural images have a 1/f^2 power spectrum, and matching this spectrum implies perceptual quality.
    Invoked in Section 4.2 and Fig. 2 to argue outputs are natural; the implication is not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Direct Ascent Synthesis: Revealing Hidden Generative Capabilities in Discriminative Models." pith.science (2026). https://pith.science/paper/MRH743WV

@misc{pith2026250207753,
  author       = {Pith},
  title        = {Pith review of: Direct Ascent Synthesis: Revealing Hidden Generative Capabilities in Discriminative Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MRH743WV}},
  note         = {Machine review of arXiv:2502.07753}
}
abstract

We demonstrate that discriminative models inherently contain powerful generative capabilities, challenging the fundamental distinction between discriminative and generative architectures. Our method, Direct Ascent Synthesis (DAS), reveals these latent capabilities through multi-resolution optimization of CLIP model representations. While traditional inversion attempts produce adversarial patterns, DAS achieves high-quality image synthesis by decomposing optimization across multiple spatial scales (1x1 to 224x224), requiring no additional training. This approach not only enables diverse applications -- from text-to-image generation to style transfer -- but maintains natural image statistics ($1/f^2$ spectrum) and guides the generation away from non-robust adversarial patterns. Our results demonstrate that standard discriminative models encode substantially richer generative knowledge than previously recognized, providing new perspectives on model interpretability and the relationship between adversarial examples and natural image synthesis.

Figures

Figures reproduced from arXiv: 2502.07753 by the authors.

Figure 1
Figure 1. Direct Ascent Synthesis generates high-quality images by optimizing multi-resolution components to match CLIP embed￾dings, without any generative training. Unlike standard adversarial optimization that produces noise-like patterns, our approach re￾veals that pretrained discriminative models contain rich generative knowledge accessible through careful optimization. It can be used for a variety of image manipulations,… view at source ↗
Figure 2
Figure 2. Multi-resolution decomposition enables training-free image synthesis. Left: An image is expressed as a sum of components at increasing resolutions, from 1 × 1 to 224 × 224. Middle: The components are optimized simultaneously to maximize CLIP embedding similarity with a target description, producing coherent images without generative training. Right: The power spectrum of generated images follows a 1/f 2 distribution… view at source ↗
Figure 3
Figure 3. Diverse generations from Direct Ascent Synthesis across a range of concepts and styles. Results were obtained by optimizing against an ensemble of three CLIP models, with prompt augmentation to control image aesthetics: discouraging text generation (-0.3 × ”Optical Character Recognition”), enhancing rendering quality (0.3 × ”octane render, unreal engine, ray tracing, volumetric lighting”), and preventing image stack… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Ablation study demonstrating how different components of Direct Ascent Synthesis contribute to coherent image generation. Left: Direct pixel optimization yields adversarial patterns typical of model inversion attacks. Middle: Adding augmentations and model ensembling b…
Figure 6
Figure 6. Figure 6: Direct Ascent Synthesis enables efficient neural style transfer without the artifacts common in pixel-space optimization. Starting from a source image and guidance image, we are able to effectively combine the two using DAS. This demonstrates that our multi￾resolution …
Figure 7
Figure 7. Figure 7: Four independent generations of ”a photo of a volcanic eruption in Iceland” on the left, and ”a beautiful photo of Cambridge UK, detailed” with an additional prompt of ”winter night”. The generations used 3 CLIP models at once, and a corrective prompt of ”Optical Chara…
Figure 9
Figure 9. Figure 9: Reconstructing an image from its embedding. Instead of a text prompt, we used an embedded original image to guide the Direct Ascent Synthesis generation. Two resulting reconstructions are shown for each image, demonstrating consistent recovery of major semantic element…
Figure 12
Figure 12. Figure 12: Applying style transfer while optimizing raw pixels (c) vs DAS (d). and generation in deep neural networks. By demonstrating that discriminative models contain rich generative knowl￾edge that can be accessed through careful optimization, we challenge several conventio…
Figure 11
Figure 11. Figure 11: An example of inpainting using Direct Ascent Synthesis. The masked image was filled in using the prompt ”a city skyline at night”, demonstrating seamless integration of generated content with existing context. dress diverse synthesis tasks without task-specific traini…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Implicit Inversion turns CLIP into a Decoder

    cs.CV 2025-05 conditional novelty 6.0 of 10

    A frozen CLIP encoder can be inverted through frequency-aware implicit neural representations to generate images, reconstruct photos, edit scenes, and transfer style, all without training a decoder.

  2. QoS-Aware Token Scheduling and Private Data Valuation for Multi-Modal Agentic Networks

    cs.AI 2026-04 conditional novelty 5.5 of 10

    DP-protected semantic prototypes plus novelty-based valuation parameterize a token-bucket AI-quota scheduler that improves contribution fairness and reconstruction resistance over baselines in COCO/CLIP simulations.

  3. Anchoring and Steering Diffusion: Enhancing the Faithfulness of Text-to-Image Generation at Inference Time

    cs.CV 2026-07 conditional novelty 5.0 of 10

    AnchorSteer improves text-to-image faithfulness by anchoring initial noise with CLIP/DAS-derived semantics (LP-SDS) and correcting errors during denoising with a VLM-driven Think-Erase-Retouch loop.

Reference graph

Works this paper leans on

39 extracted references · 19 canonical work pages · cited by 3 Pith papers

  1. [1]

    write newline

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

  2. [2]

    Gan dissection: Visualizing and understanding generative adversarial networks

    Bau, D., Zhu, J.-Y., Wulff, J., Peebles, W., Strobelt, H., Zhou, B., and Torralba, A. Gan dissection: Visualizing and understanding generative adversarial networks. In International Conference on Learning Representations, 2019 a

  3. [3]

    Seeing What a GAN Cannot Generate

    Bau, D., Zhu, J.-Y., Wulff, J., Peebles, W., Strobelt, H., Zhou, B., and Torralba, A. Seeing what a gan cannot generate, 2019 b . URL https://arxiv.org/abs/1910.11626

  4. [4]

    and Adelson, E

    Burt, P. and Adelson, E. The laplacian pyramid as a compact image code. IEEE Transactions on Communications, 31 0 (4): 0 532--540, 1983. doi:10.1109/TCOM.1983.1095851

  5. [5]

    Reproducible scaling laws for contrastive language-image learning

    Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J. Reproducible scaling laws for contrastive language-image learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 2818–2829. IEEE, June 2023. doi:10.1109/cvpr52729.2023.00276. URL http://dx.doi.org/...

  6. [6]

    Vqgan-clip: Open domain image generation and editing with natural language guidance

    Crowson, K., Biderman, S., Kornis, D., Stander, D., Hallahan, E., Castricato, L., and Raff, E. Vqgan-clip: Open domain image generation and editing with natural language guidance. In European Conference on Computer Vision, pp.\ 88--105. Springer, 2022

  7. [7]

    F., Mahendran, A., Yu, F., Oliver, A., Huot, F., Bastings, J., Collier, M

    Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A., Caron, M., Geirhos, R., Alabdulmohsin, I., Jenatton, R., Beyer, L., Tschannen, M., Arnab, A., Wang, X., Riquelme, C., Minderer, M., Puigcerver, J., Evci, U., Kumar, M., van Steenkiste, S., Elsayed, G. F., Mahendran, A., Yu, F., Oliver, A., Huot, F., Bastings, J., Co...

  8. [8]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. URL https://arxiv.org/abs/2010.11929

Show all 39 references
  1. [9]

    A mathematical framework for transformer circuits

    Elhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., Askell, A., Ndousse, K., Amodei, D., Jones, A., DasSarma, N., Askell, D., Wang, P., and Chen, A. A mathematical framework for transformer circuits. Anthropic Technical Report, 2022

  2. [10]

    Adversarial examples for the openai clip in its zero-shot classification regime and their semantic generalization, Jan 2021 a

    Fort, S. Adversarial examples for the openai clip in its zero-shot classification regime and their semantic generalization, Jan 2021 a . URL https://stanislavfort.github.io/2021/01/12/OpenAI_CLIP_adversarial_examples.html

  3. [11]

    Pixels still beat text: Attacking the openai clip model with text patches and adversarial pixel perturbations, March 2021 b

    Fort, S. Pixels still beat text: Attacking the openai clip model with text patches and adversarial pixel perturbations, March 2021 b . URL https://stanislavfort.github.io/2021/03/05/OpenAI_CLIP_stickers_and_adversarial_examples.html

  4. [12]

    Adversarial vulnerability of powerful near out-of-distribution detection, 2022

    Fort, S. Adversarial vulnerability of powerful near out-of-distribution detection, 2022. URL https://arxiv.org/abs/2201.07012

  5. [13]

    A note on implementation errors in recent adaptive attacks against multi-resolution self-ensembles, 2025

    Fort, S. A note on implementation errors in recent adaptive attacks against multi-resolution self-ensembles, 2025. URL https://arxiv.org/abs/2501.14496

  6. [14]

    and Lakshminarayanan, B

    Fort, S. and Lakshminarayanan, B. Ensemble everything everywhere: Multi-scale aggregation for adversarial robustness, 2024. URL https://arxiv.org/abs/2408.05446

  7. [15]

    Exploring the limits of out-of-distribution detection, 2021

    Fort, S., Ren, J., and Lakshminarayanan, B. Exploring the limits of out-of-distribution detection, 2021. URL https://arxiv.org/abs/2106.03004

  8. [16]

    D., Ganguli, S., and Schoenholz, S

    Fort, S., Cubuk, E. D., Ganguli, S., and Schoenholz, S. S. What does a deep neural network confidently perceive? the effective dimension of high certainty class manifolds and their low confidence boundaries, 2022. URL https://arxiv.org/abs/2210.05546

  9. [17]

    A., Ecker, A

    Gatys, L. A., Ecker, A. S., and Bethge, M. Image style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016

  10. [18]

    J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y

    Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial networks, 2014. URL https://arxiv.org/abs/1406.2661

  11. [19]

    J., Shlens, J., and Szegedy, C

    Goodfellow, I. J., Shlens, J., and Szegedy, C. Explaining and harnessing adversarial examples, 2015. URL https://arxiv.org/abs/1412.6572

  12. [20]

    Denoising diffusion probabilistic models, 2020

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models, 2020. URL https://arxiv.org/abs/2006.11239

  13. [21]

    Hyv \"a rinen, A., Hurri, J., and Hoyer, P. O. Natural Image Statistics: A Probabilistic Approach to Early Computational Vision, volume 39 of Computational Imaging and Vision. Springer, 2009. ISBN 978-1-84882-491-1. doi:10.1007/978-1-84882-491-1

  14. [22]

    Adversarial examples are not bugs, they are features, 2019

    Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., and Madry, A. Adversarial examples are not bugs, they are features, 2019. URL https://arxiv.org/abs/1905.02175

  15. [23]

    Factors influencing spatial frequency extraction in faces: A review

    Jeantet, C., Caharel, S., Schwan, R., Lighezzolo-Alnot, J., and Laprevote, V. Factors influencing spatial frequency extraction in faces: A review. Neuroscience and Biobehavioral Reviews, 93: 0 123--138, 2018. ISSN 0149-7634. doi:https://doi.org/10.1016/j.neubiorev.2018.03.006....

  16. [24]

    and Ganguli, S

    Kamb, M. and Ganguli, S. An analytic theory of creativity in convolutional diffusion models, 2024. URL https://arxiv.org/abs/2412.20292

  17. [25]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes, 2022. URL https://arxiv.org/abs/1312.6114

  18. [26]

    Scale-space theory: a basic tool for analyzing structures at different scales

    Lindeberg, T. Scale-space theory: a basic tool for analyzing structures at different scales. Journal of Applied Statistics, 21 0 (1-2): 0 225--270, 1994. doi:10.1080/757582976. URL https://doi.org/10.1080/757582976

  19. [27]

    Towards deep learning models resistant to adversarial attacks, 2019

    Madry, A., Makelov, A., Schmidt, L., Tsipras, D., and Vladu, A. Towards deep learning models resistant to adversarial attacks, 2019. URL https://arxiv.org/abs/1706.06083

  20. [28]

    and Vedaldi, A

    Mahendran, A. and Vedaldi, A. Understanding deep image representations by inverting them, 2014

  21. [29]

    Synthesizing the preferred inputs for neurons in neural networks via deep generator networks, 2016

    Nguyen, A., Dosovitskiy, A., Yosinski, J., Brox, T., and Clune, J. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks, 2016. URL https://arxiv.org/abs/1605.09304

  22. [30]

    Feature visualization

    Olah, C., Mordvintsev, A., and Schubert, L. Feature visualization. Distill, 2017. doi:10.23915/distill.00007. https://distill.pub/2017/feature-visualization

  23. [31]

    Zoom in: An introduction to circuits

    Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M., and Carter, S. Zoom in: An introduction to circuits. Distill, 2020. doi:10.23915/distill.00024.001

  24. [32]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision, 2021. URL https://arxiv.org/abs/2103.00020

  25. [33]

    and Monro, S

    Robbins, H. and Monro, S. A stochastic approximation method. The Annals of Mathematical Statistics, 22 0 (3): 0 400--407, September 1951. doi:10.1214/aoms/1177729586. URL https://doi.org/10.1214

  26. [34]

    High-resolution image synthesis with latent diffusion models, 2022

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models, 2022. URL https://arxiv.org/abs/2112.10752

  27. [35]

    Ruderman, D. L. Statistics of natural images: Scaling analysis and the scale-space paradigm. Physical Review Letters, 73 0 (6): 0 814--817, 1994. doi:10.1103/PhysRevLett.73.814

  28. [36]

    Deep image prior

    Ulyanov, D., Vedaldi, A., and Lempitsky, V. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

  29. [37]

    and van Hateren , J

    van der Schaaf , A. and van Hateren , J. Modelling the power spectra of natural images: Statistics and information. Vision Research, 36 0 (17): 0 2759--2770, September 1996. ISSN 0042-6989. Relation: http://www.rug.nl/informatica/organisatie/overorganisatie/iwi Rights: Univers...

  30. [38]

    imstack: Image stack exploration and analysis

    Whitaker, J. imstack: Image stack exploration and analysis. https://johnowhitaker.github.io/imstack/, 2022. Accessed: 2024-01-31

  31. [39]

    Understanding neural networks through deep visualization, 2015

    Yosinski, J., Clune, J., Nguyen, A., Fuchs, T., and Lipson, H. Understanding neural networks through deep visualization, 2015. URL https://arxiv.org/abs/1506.06579

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.