Pith. sign in

REVIEW 3 major objections 5 minor 66 references

Diffusion-based Visual Anagram as Multi-task Learning

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

Pith's one-line read Casting visual anagram generation as multi-task learning, with anti-segregation attention overlap, completion-based noise reweighting, and variance rectification, produces anagrams that avoid concept segregation and domination.

desk verdict Solid incremental paper on visual anagrams: new multi-task framing, clean math, but the NVR derivation rests on an unverified Gaussian-noise assumption and the empirical evidence lacks error bars. read the letter →

arxiv 2412.02693 v1 pith:S2O63L76 submitted 2024-12-03 cs.CV

classification cs.CV
keywords visualanagramsdiffusionmodelsmulti-tasklearningopticalillusionscross-attentionmapsnoisevectorbalancingvariancerectificationtext-to-imagegeneration
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

This paper claims that visual anagram generation with diffusion models fails when noise predictions from different views are simply averaged, because the concepts split apart or one dominates. It reframes the problem as multi-task learning: each view is a task, and the denoising trajectory should be a compromise that serves all tasks at once. The authors introduce three interventions: encouraging cross-view overlap in attention maps, reweighting each view's noise by how close that view already is to completion, and rescaling the combined noise to restore the variance the diffusion model expects. On two-view and three-view benchmarks they report higher worst-view alignment, better concealment, and better average text-image alignment than the averaging baseline, without training or fine-tuning the diffusion model.

What carries the argument

Three mechanisms carry the argument. Anti-segregation optimization operates on the cross-attention maps of the subject tokens in each view and minimizes the difference between a target overlap ratio and the actual intersection-over-union of those maps, applying a one-step gradient modulation to the denoised image. Noise-vector balancing measures each view's task completion as the cosine similarity between a text embedding and a noise-aware image embedding of the currently denoised image, then reweights that view's predicted noise inversely to its completion with a time-dependent exponent. Noise-variance rectification treats the combined noise as a weighted sum of per-view noise vectors assumed to be standard normal with a constant per-element correlation, estimates that correlation from the dot product of the two noise maps, and rescales the combined noise by the single factor that restores unit variance.

What would settle it

Record the per-element variances and pairwise correlations of the predicted noise vectors from the pixel-space diffusion model across prompts and timesteps; if the variances are not close to one or the correlation varies strongly across spatial locations, the single scale factor in the rectification equation cannot restore unit variance, and the rectification step rests on a false premise. A simpler check: replace the rectified noise with an empirical re-standardization per channel and see whether image quality changes.

Watch

Extended reading notes

Core claim

The central claim is that the two failure modes of prior visual anagram methods, concept segregation and concept domination, are symptoms of treating each view's generation as independent, and that a multi-task learning view of the denoising trajectory fixes both. The paper's contributions are the anti-segregation loss, which pushes the spatial attention maps of different prompts' subject tokens to overlap in the shared image; the noise-vector balancing rule, which down-weights views whose task-completion score is already high; and the noise-variance rectification, which multiplies the averaged noise by a scale factor derived from the estimated correlation between views so the combined noise keeps unit variance. In the paper's own terms, the result is a denoising trajectory that aligns well across tasks, generating images that read as one concept under one view and a different concept under the transformed view rather than two objects pasted into one frame.

Load-bearing premise

The rectification step assumes that the diffusion model's predicted noise for every view is standard Gaussian and that the correlation between views is the same for every pixel, so a single multiplicative constant can restore the correct statistics; if the predicted noise is not standard Gaussian, the correction can itself distort the denoising trajectory.

Editorial extensions

If this is right

  • On the paper's two-view CIFAR-10-style benchmark, the full method raises worst-view alignment from 0.2583 to 0.2711 and concealment from 0.6744 to 0.6913 relative to the averaging baseline.
  • The same gains hold on three-view prompts, indicating that the balancing and rectification machinery scales beyond pairs of views.
  • Each of the three components contributes independently, and the best results come from their combination, so the multi-task framing is doing real work rather than one trick carrying the method.
  • Because the pipeline needs no training or fine-tuning, the method can be plugged into any pixel-space diffusion model that exposes cross-attention maps and per-view noise predictions.

Reading between the lines

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

  • The variance-rectification step is generic: any method that averages or interpolates noise predictions from multiple prompts, not just anagram generation, could inherit unit-variance corrections, and its benefit could be tested on ordinary multi-concept composition.
  • The target overlap ratio is a hyperparameter tuned per dataset; one could make it prompt-adaptive, raising it for concepts that naturally share silhouette structure and lowering it for unrelated pairs, which the paper does not explore.
  • The correlation coefficients the paper measures are close to one across timesteps, which suggests the main practical role of rectification is protective rather than large-magnitude; ablations show it mostly lifts average alignment, so its effect may be on image quality rather than on concept blending.
  • If the predicted noise from a diffusion model is not actually standard normal, the single scalar correction will not restore exact statistics; checking the marginal distribution of the predicted noise would resolve whether the derivation's premise holds.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper addresses two failure modes in diffusion-based visual anagram generation: concept segregation (different concepts are generated independently rather than sharing a single image) and concept domination (one concept overwhelms the others). It casts anagram generation as a multi-task learning problem and contributes three inference-time techniques: Anti-Segregation Optimization (ASO), which encourages overlap between cross-attention maps of different views via a loss term; Noise Vector Balancing (NVB), which reweights each view's predicted noise by a CLIP-based task completion score; and Noise Variance Rectification (NVR), which scales the combined noise to restore unit variance under an assumed Gaussian noise model. Experiments on 2-view and 3-view CIFAR-10 prompts, plus free-form examples, compare against the previous baseline (Geng et al., 2024) and report consistent improvements in worst alignment, concealment, and average alignment, with an ablation study attributing gains to each component.

Significance. If the reported improvements are robust, the paper offers a practical, training-free way to improve visual anagram quality using only off-the-shelf diffusion and CLIP models, which is useful for a growing community working on diffusion-based illusions. The multi-task learning analogy is clearly presented, and the ablations show that each proposed component contributes positively to at least one metric. The paper also ships code, which supports reproducibility. However, the quantitative gains are small (on the order of 0.01 in cosine-similarity-based metrics), error bars are absent, and at least one hyperparameter (phi) appears to be tuned on the evaluation set, so the significance of the headline improvements is currently uncertain.

major comments (3)
  1. [Sec. 3.4, Eqs. (7)-(12)] The NVR derivation is built on the assumption in Eq. (7) that each predicted noise vector epsilon_i^t is standard Gaussian. In practice, epsilon_theta(x_t, t, y_i) is a deterministic function of the current noisy image and the prompt, and its marginal distribution may differ from N(0, I). The paper never verifies this assumption, and Fig. 10 only reports cosine similarity and the correlation coefficient of Eq. (12), not marginal variances. If the variance of the predicted noises is not 1, the scalar c computed in Eq. (10) will not set the combined noise variance to 1, so the claimed statistical rectification mechanism is unsupported. Please add an empirical check of the marginal variance and distribution of predicted noises across views and timesteps, and verify that the rectified noise indeed has unit variance; if it does not, the explanation for the NVR improvement in Table 2 would need to be revised.
  2. [Sec. 5.2 and Fig. 9; Table 2] The target overlap hyperparameter phi is selected on the same 2-view CIFAR-10 evaluation set that is used for the headline comparisons: Fig. 9 reports that the best performance is achieved around phi=0.45, and the main tables presumably use this value. This gives the proposed method an advantage over the baseline, which has no equivalent evaluation-set tuning, and makes the reported gains optimistic. Please either fix phi before evaluation, tune it on a separate validation split, or report the main results across a range of phi values so readers can assess sensitivity.
  3. [Table 2 and Sec. 4.2] All metrics are averages over 10 seeds per prompt pair, but no standard deviations, confidence intervals, or per-pair breakdowns are provided. The reported improvements are small: for example, Amin rises from 0.2583 to 0.2711 and Aavg from 0.2717 to 0.2816, while several ablation rows differ by less than 0.01 in C. Without error bars or a paired significance test over the 10 seeds, it is difficult to judge whether the claimed 'superior ability' and the component-wise contributions are statistically reliable. Please include error bars or a bootstrap test.
minor comments (5)
  1. [Sec. 1, Contributions] There is a typo in the third contribution: 'satet-of-the-art' should be 'state-of-the-art'.
  2. [Eq. (1) and Sec. 5.2] The loss in Eq. (1) penalizes absolute deviation from the target overlap phi, so it does not monotonically push toward larger overlap for ratios above phi. The intuitive description in Sec. 5.2 ('a larger phi encourages more overlap') is acceptable, but the exact behavior near the target could be clarified in the text.
  3. [Sec. 3.3] The noise-aware CLIP image encoder is taken from GLIDE [39], but it is not stated whether this model has the same architecture and embedding space as the vanilla CLIP used for evaluation. Please specify the relation between the two CLIP models for reproducibility.
  4. [Appendix A and Fig. 11] Score Distillation Loss appears as '[?]' in the appendix text and figure caption; it should be a proper citation, e.g., to Poole et al. [43] which is already in the reference list.
  5. [Eq. (10)] The formula for c in Eq. (10) is rendered with an unclear square-root symbol; please ensure the mathematical typesetting is unambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the NVR derivation is an explicit variance-normalization identity, and the evaluation metrics are not optimized by construction.

full rationale

I inspected the claimed derivation chain. The only first-principles-style derivation is Noise Variance Rectification (Sec. 3.4), which starts from the explicit assumption in Eq. (7) that each per-view predicted noise vector is standard Gaussian, then computes the scale factor c that makes the combined noise variance equal to one. This is an algebraic normalization, not a prediction of a quantity that is then verified from the same quantity; the scale factor is defined as the variance-preserving factor, so its effect is a direct consequence of the equation, not a hidden fit. The Gaussian assumption itself is unverified and may be questioned as a modeling assumption, but that is a correctness/robustness concern, not circularity. The anti-segregation objective (Eq. (1)) is explicitly introduced as a conjecture about attention-map overlap, and it does not directly optimize any reported metric; the reported CLIP-based metrics are computed with a different, vanilla CLIP model than the noise-aware CLIP used for task-completion weighting, and the paper states this explicitly. Hyperparameter phi is tuned on the 2-view evaluation set (Sec. 5.2), which is a potential selection issue, but it is not a fitted parameter renamed as a prediction. Self-citations (e.g., [8], [9], [10], [17], [31], [60], [63]) appear only in background enumerations of diffusion models and multi-task learning and are not load-bearing for any central claim. No step in the paper reduces by construction to its own inputs or to a self-citation chain.

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

The method introduces no new entities, but it relies on several assumptions: Gaussian predicted noise, constant cross-view correlation, a single identifiable concept per prompt, and CLIP as a task-completion oracle. The free parameters (phi, NVB exponent and slope, ASO step size) are either tuned on the evaluation set or chosen without ablation.

free parameters (4)
  • phi (target overlap ratio in ASO) = 0.45
    Controls the desired overlap between cross-attention maps of different views in Eq. (1); tuned on the 2-view CIFAR10 dataset in Fig. 9, with best value around 0.45. Used in all final experiments.
  • cosine similarity exponent in NVB = -2
    The negative exponent in Eq. (4) reweights noise vectors by task completion; the value -2 is chosen without derivation or ablation.
  • t/T linear term coefficient in NVB = 1
    The linear schedule in Eq. (4) increases the influence of task completion as denoising progresses; the slope 1 is an ad hoc choice not ablated.
  • ASO gradient step size alpha = not reported
    Controls the strength of anti-segregation optimization in Eq. (2); no value or sensitivity analysis is given in the paper.
assumptions (4)
  • domain assumption Predicted noise vectors from each view are standard Gaussian (Eq. 7)
    Used to derive the NVR scale factor; diffusion model outputs are not guaranteed to have unit variance.
  • ad hoc to paper Correlation coefficients between noise vectors of different views are constant across all elements within a denoising step
    Needed to estimate a single covariance per view pair via the law of large numbers (Eq. 12); an approximation that may not hold.
  • domain assumption Each prompt contains exactly one concept that can be identified by its tokens
    The ASO loss requires selecting the main object tokens in each prompt; this limits the method to simple prompts.
  • domain assumption CLIP cosine similarity between text and a noised view of the image is a valid task completion signal
    Used in NVB to reweight noise; relies on the noise-aware CLIP model providing meaningful scores on intermediate latents.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diffusion-based Visual Anagram as Multi-task Learning." pith.science (2026). https://pith.science/paper/S2O63L76

@misc{pith2026241202693,
  author       = {Pith},
  title        = {Pith review of: Diffusion-based Visual Anagram as Multi-task Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S2O63L76}},
  note         = {Machine review of arXiv:2412.02693}
}
read the original abstract

Visual anagrams are images that change appearance upon transformation, like flipping or rotation. With the advent of diffusion models, generating such optical illusions can be achieved by averaging noise across multiple views during the reverse denoising process. However, we observe two critical failure modes in this approach: (i) concept segregation, where concepts in different views are independently generated, which can not be considered a true anagram, and (ii) concept domination, where certain concepts overpower others. In this work, we cast the visual anagram generation problem in a multi-task learning setting, where different viewpoint prompts are analogous to different tasks,and derive denoising trajectories that align well across tasks simultaneously. At the core of our designed framework are two newly introduced techniques, where (i) an anti-segregation optimization strategy that promotes overlap in cross-attention maps between different concepts, and (ii) a noise vector balancing method that adaptively adjusts the influence of different tasks. Additionally, we observe that directly averaging noise predictions yields suboptimal performance because statistical properties may not be preserved, prompting us to derive a noise variance rectification method. Extensive qualitative and quantitative experiments demonstrate our method's superior ability to generate visual anagrams spanning diverse concepts.

Figures

Figures reproduced from arXiv: 2412.02693 by the authors.

Figure 1
Figure 1. Visual Anagrams. We show an example of visual ana￾grams, which can be perceived as a garden gnome or a hot air balloon depending on the orientation of the image. a rabbit a violin (a) a street art of ... a cat a ship (b) a painting of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Failure Cases. We show two common failure cases of [19]: concept segregation (left) and domination (right). Diffusion. With the advent of diffusion models [14, 26, 50, 52, 60], generating such visual anagrams has become significantly easier by simply averaging noise predictions across different views of various prompts. In the reverse denoising process of diffusion models, the image layout is formed during the early… view at source ↗
Figure 3
Figure 3. Method overview. During each denoising step, the intermediate image xt first passes through the diffusion model together with the corresponding text prompt under each view, and also through a noise-aware CLIP model which measures the degree of task completion for each view. (1) Noise Vector Balancing: Predicted noise vectors are reweighted based on the degree of task completion before being combined, see Sec. 3.3. (… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: More qualitative results of our proposed method compared to the baseline method [ [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results of Anti-Segregation Optimization and [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results of Noise Variance Rectification. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 9
Figure 9. Figure 9: Effect of target overlap ratio in Anti-Segregation Opti [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: Correlation between noise vectors. We record the cosine similarity and estimated correlation coefficient between the predicted noise vectors of different views at each denoising step. and now the upper-right background region in the flipped view naturally interprets a…
Figure 11
Figure 11. Figure 11: Qualitative Results. We provide additional qualitative results to compare our method with existing methods. Tancik [54] uses a latent diffusion model [45] but struggles with transformation inconsistencies in the latent code, as discussed in [19]. Burgert et al. [5] em…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

66 extracted references · 47 canonical work pages

  1. [1]

    https://stability.ai/news/deepfloyd-if-text-to-image- model

    Stability ai releases deepfloyd if, a powerful text-to- image model that can smartly integrate text into im- ages. https://stability.ai/news/deepfloyd-if-text-to-image- model. 2, 6

  2. [2]

    Aishwarya Agarwal, Srikrishna Karanam, K. J. Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srini- vasan. A-star: Test-time attention segregation and retention for text-to-image synthesis, 2023. 2

  3. [3]

    Separate-and-enhance: Compo- sitional finetuning for text2image diffusion models, 2024

    Zhipeng Bao, Yijun Li, Krishna Kumar Singh, Yu-Xiong Wang, and Martial Hebert. Separate-and-enhance: Compo- sitional finetuning for text2image diffusion models, 2024. 2

  4. [4]

    A new ambiguous figure

    Edwin G Boring. A new ambiguous figure. The American Journal of Psychology, 1930. 1

  5. [5]

    Diffusion Illusions: Hiding Images in Plain Sight

    Ryan Burgert, Xiang Li, Abe Leite, Kanchana Ranasinghe, and Michael Ryoo. Diffusion Illusions: Hiding Images in Plain Sight. In ACM SIGGRAPH 2024 Conference Papers, SIGGRAPH ’24, pages 1–11, New York, NY , USA, July

  6. [6]

    Designing perceptual puzzles by differentiating probabilistic programs

    Kartik Chandra, Tzu-Mao Li, Joshua Tenenbaum, and Jonathan Ragan-Kelley. Designing perceptual puzzles by differentiating probabilistic programs. In Special Interest Group on Computer Graphics and Interactive Techniques Conference Proceedings, pages 1–9, 2022. 2

  7. [7]

    Attend-and-excite: Attention-based se- mantic guidance for text-to-image diffusion models, 2023

    Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based se- mantic guidance for text-to-image diffusion models, 2023. 2

  8. [8]

    Ultraman: Single image 3d human reconstruction with ultra speed and detail

    Mingjin Chen, Junhao Chen, Xiaojun Ye, Huan-ang Gao, Xi- aoxue Chen, Zhaoxin Fan, and Hao Zhao. Ultraman: Single image 3d human reconstruction with ultra speed and detail. arXiv preprint arXiv:2403.12028, 2024. 2

Show all 66 references
  1. [9]

    Cerberus transformer: Joint semantic, affor- dance and attribute parsing

    Xiaoxue Chen, Tianyu Liu, Hao Zhao, Guyue Zhou, and Ya-Qin Zhang. Cerberus transformer: Joint semantic, affor- dance and attribute parsing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 19649–19658, 2022. 3

  2. [10]

    Pq-transformer: Jointly parsing 3d objects and layouts from point clouds

    Xiaoxue Chen, Hao Zhao, Guyue Zhou, and Ya-Qin Zhang. Pq-transformer: Jointly parsing 3d objects and layouts from point clouds. IEEE Robotics and Automation Letters , 7(2):2519–2526, 2022. 3

  3. [11]

    Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks, 2018

    Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and An- drew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks, 2018. 3, 4

  4. [12]

    Just pick a sign: Optimizing deep multitask models with gradient sign dropout

    Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. https://arxiv.org/abs/2010.06808v1,

  5. [13]

    Mitra, Daniel Cohen- Or, Tien-Tsin Wong, and Tong-Yee Lee

    Hung Kuo Chu, Wei Hsin Hsu, Niloy J. Mitra, Daniel Cohen- Or, Tien-Tsin Wong, and Tong-Yee Lee. Camouflage im- ages. ACM Transactions on Graphics, 29(4):51, 2010. 2

  6. [14]

    Diffu- sion models beat gans on image synthesis

    Prafulla Dhariwal and Alex Nichol. Diffu- sion models beat gans on image synthesis. https://arxiv.org/abs/2105.05233v4, 2021. 1, 2

  7. [15]

    Arcimboldo

    Sylvia Ferino-Pagden. Arcimboldo. Skira, 2017. 1

  8. [16]

    Freeman, Edward H

    William T. Freeman, Edward H. Adelson, and David J. Heeger. Motion without movement. ACM SIGGRAPH Com- puter Graphics, 25(4):27–30, 1991. 2

  9. [17]

    Scp-diff: Photo-realistic semantic image synthesis with spatial-categorical joint prior

    Huan-ang Gao, Mingju Gao, Jiaju Li, Wenyi Li, Rong Zhi, Hao Tang, and Hao Zhao. Scp-diff: Photo-realistic semantic image synthesis with spatial-categorical joint prior. arXiv preprint arXiv:2403.09638, 2024. 2

  10. [18]

    Visual ana- grams: Generating multi-view optical illusions with diffu- sion models

    Daniel Geng, Inbum Park, and Andrew Owens. Visual ana- grams: Generating multi-view optical illusions with diffu- sion models. https://arxiv.org/abs/2311.17919v2, 2023. 2, 6

  11. [19]

    Visual ana- grams: Generating multi-view optical illusions with diffu- sion models

    Daniel Geng, Inbum Park, and Andrew Owens. Visual ana- grams: Generating multi-view optical illusions with diffu- sion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24154– 24163, 2024. 1, 2, 3, 4, 6, 8, 11, 12

  12. [20]

    Convolutional neural networks can be deceived by visual illusions

    Alexander Gomez-Villa, Adrian Martin, Javier Vazquez- Corral, and Marcelo Bertalmio. Convolutional neural networks can be deceived by visual illusions. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12301–12309, Long Beach, CA, USA, 2019. 1, 2

  13. [21]

    Diffusion models as plug-and-play priors,

    Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors,

  14. [22]

    Semantically-guided represen- tation learning for self-supervised monocular depth

    Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided represen- tation learning for self-supervised monocular depth. https://arxiv.org/abs/2002.12319v1, 2020. 3

  15. [23]

    Dynamic task prioritization for multitask learning

    Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, and Li Fei-Fei. Dynamic task prioritization for multitask learning. In Vittorio Ferrari, Martial Hebert, Cristian Smin- chisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 282–299, Cham, 2018. 3

  16. [24]

    Ganmouflage: 3d object nondetection with texture fields, 2023

    Rui Guo, Jasmine Collins, Oscar de Lima, and Andrew Owens. Ganmouflage: 3d object nondetection with texture fields, 2023. 2

  17. [25]

    Visual indeterminacy in gan art

    Aaron Hertzmann. Visual indeterminacy in gan art. Leonardo, 53(4):424–428, 2020. 2

  18. [26]

    Denoising diffu- sion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models, 2020. 1, 2, 5

  19. [27]

    Multi-task curriculum learning based on gradient similarity

    Hiroaki Igarashi. Multi-task curriculum learning based on gradient similarity. BMVC, 2022. 3

  20. [28]

    Intriguing properties of generative classifiers, 2024

    Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers, 2024. 1, 2

  21. [29]

    Multi-task learning using uncertainty to weigh losses for scene geome- try and semantics, 2018

    Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geome- try and semantics, 2018. 3

  22. [30]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 6

  23. [31]

    Fairdiff: Fair segmenta- tion with point-image diffusion

    Wenyi Li, Haoran Xu, Guiyu Zhang, Huan-ang Gao, Mingju Gao, Mengyu Wang, and Hao Zhao. Fairdiff: Fair segmenta- tion with point-image diffusion. In International Conference on Medical Image Computing and Computer-Assisted Inter- vention, pages 617–628. Springer, 2024. 2

  24. [32]

    Pareto multi-task learning, 2019

    Xi Lin, Hui-Ling Zhen, Zhenhua Li, Qingfu Zhang, and Sam Kwong. Pareto multi-task learning, 2019. 3

  25. [33]

    Tenenbaum

    Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional vi- sual generation with composable diffusion models. https://arxiv.org/abs/2206.01714v6, 2022. 2

  26. [34]

    A Closed-Form Solution to Universal Style Transfer

    Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A Closed-Form Solution to Universal Style Transfer. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV) , pages 5951–5960, Seoul, Korea (South), Oct. 2019. IEEE. 2

  27. [35]

    Decoder network over lightweight recon- structed feature for fast semantic style transfer

    Ming Lu, Hao Zhao, Anbang Yao, Feng Xu, Yurong Chen, and Li Zhang. Decoder network over lightweight recon- structed feature for fast semantic style transfer. In Proceed- ings of the IEEE international conference on computer vi- sion, pages 2469–2477, 2017. 2

  28. [36]

    Cross-stitch networks for multi-task learning,

    Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Mar- tial Hebert. Cross-stitch networks for multi-task learning,

  29. [37]

    Multi-task learning as a bargaining game

    Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multi-task learning as a bargaining game. https://arxiv.org/abs/2202.01017v2, 2022. 3

  30. [38]

    Sankaranarayanan, and Phillip Isola

    Jerry Ngo, S. Sankaranarayanan, and Phillip Isola. Is clip fooled by optical illusions? In Tiny Papers @ ICLR, 2023. 1, 2

  31. [39]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022

    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, 2022. 6

  32. [40]

    Michael E. R. Nicholls, Owen Churches, and Tobias Loetscher. Perception of an ambiguous figure is affected by own-age social biases. Scientific Reports, 8(1):12661, Aug

  33. [41]

    Aude Oliva, Antonio Torralba, and Philippe G. Schyns. Hy- brid images. ACM Trans. Graph., 25(3):527–532, 2006. 2

  34. [42]

    Camouflaging an object from many viewpoints

    Andrew Owens, Connelly Barnes, Alex Flint, Hanumant Singh, and William Freeman. Camouflaging an object from many viewpoints. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2782–2789, 2014. 2

  35. [43]

    Dreamfusion: Text-to-3d using 2d diffusion

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2

  36. [44]

    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. https://arxiv.org/abs/2102.12092v2, 2021. 2

  37. [45]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High- resolution image synthesis with latent diffusion models. https://arxiv.org/abs/2112.10752v2, 2021. 2, 3, 11, 12

  38. [46]

    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...

  39. [47]

    Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion mo...

  40. [48]

    Multi-task learning as multi-objective optimization, 2019

    Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization, 2019. 3

  41. [49]

    Independent component alignment for multi-task learning, 2023

    Dmitry Senushkin, Nikolay Patakin, Arseny Kuznetsov, and Anton Konushin. Independent component alignment for multi-task learning, 2023. 3

  42. [50]

    Weiss, Niru Mah- eswaranathan, and Surya Ganguli

    Jascha Sohl-Dickstein, Eric A. Weiss, Niru Mah- eswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. https://arxiv.org/abs/1503.03585v8, 2015. 1, 2

  43. [51]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Er- mon. Denoising diffusion implicit models. https://arxiv.org/abs/2010.02502v4, 2020. 2

  44. [52]

    Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. https://arxiv.org/abs/2011.13456v2, 2020. 1, 2

  45. [53]

    Adashare: Learning what to share for efficient deep multi-task learning

    Ximeng Sun, Rameswar Panda, Rogerio Feris, and Kate Saenko. Adashare: Learning what to share for efficient deep multi-task learning. In H. Larochelle, M. Ranzato, R. Had- sell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems , volume 33, pages 8728–

  46. [54]

    tancik/Illusion-Diffusion, July 2024

    Matthew Tancik. tancik/Illusion-Diffusion, July 2024. original-date: 2023-02-12T22:39:28Z. 2, 6, 11, 12

  47. [55]

    Attention is all you need

    A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3

  48. [56]

    Pretraining is all you need for image-to-image translation, 2022

    Tengfei Wang, Ting Zhang, Bo Zhang, Hao Ouyang, Dong Chen, Qifeng Chen, and Fang Wen. Pretraining is all you need for image-to-image translation, 2022. 5

  49. [57]

    Toward quantifying ambiguities in artistic images,

    Xi Wang, Zoya Bylinskii, Aaron Hertzmann, and Robert Pepperell. Toward quantifying ambiguities in artistic images,

  50. [58]

    Do current multi-task optimization meth- ods in deep learning even help?, 2022

    Derrick Xin, Behrooz Ghorbani, Ankush Garg, Orhan Firat, and Justin Gilmer. Do current multi-task optimization meth- ods in deep learning even help?, 2022. 3

  51. [59]

    Gradient surgery for multi-task learning, 2020

    Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning, 2020. 3, 4

  52. [60]

    Ctrl-u: Robust conditional image genera- tion via uncertainty-aware reward modeling

    Guiyu Zhang, Huan-ang Gao, Zijian Jiang, Hao Zhao, and Zhedong Zheng. Ctrl-u: Robust conditional image genera- tion via uncertainty-aware reward modeling. arXiv preprint arXiv:2410.11236, 2024. 1

  53. [61]

    An overview of multi-task learn- ing

    Yu Zhang and Qiang Yang. An overview of multi-task learn- ing. National Science Review, 5(1):30–43, 09 2017. 3

  54. [62]

    A survey on multi-task learn- ing

    Yu Zhang and Qiang Yang. A survey on multi-task learn- ing. IEEE transactions on knowledge and data engineering, 34(12):5586–5609, 2021. 3

  55. [63]

    Steps: Joint self-supervised nighttime image enhancement and depth estimation

    Yupeng Zheng, Chengliang Zhong, Pengfei Li, Huan-ang Gao, Yuhang Zheng, Bu Jin, Ling Wang, Hao Zhao, Guyue Zhou, Qichao Zhang, et al. Steps: Joint self-supervised nighttime image enhancement and depth estimation. In 2023 IEEE International Conference on Robotics and Automation...

  56. [2018]

    Publisher: Nature Publishing Group. 1

  57. [2024]

    1, 2, 6, 11, 12

    Association for Computing Machinery. 1, 2, 6, 11, 12

  58. [8740]

    Curran Associates, Inc., 2020. 3

Pith tools

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