Pith. sign in

REVIEW 4 major objections 6 minor 62 references

Generate, Refine, and Encode: Leveraging Synthesized Novel Samples for On-the-Fly Fine-Grained Category Discovery

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read DiffGRE synthesizes virtual-category images by interpolating between known-category images in diffusion and CLIP latent spaces, and these synthetic samples consistently improve on-the-fly fine-grained category discovery.

desk verdict A well-built plug-and-play augmentation for on-the-fly category discovery; the gains look real, but the paper's own mechanism story is only qualitatively supported. read the letter →

arxiv 2507.04051 v1 pith:GJWXRYDO submitted 2025-07-05 cs.CV

Xiao Liu , Nan Pu , Haiyang Zheng , Wenjing Li , Nicu Sebe , Zhun Zhong This is my paper
classification cs.CV
keywords On-the-flyCategoryDiscoveryfine-graineddiffusion-baseddatasynthesisattributecompositiongenerationvirtualsemi-supervisedleaderencodingonlineclusteringinferenceplug-and-playframework
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 proposes DiffGRE, a plug-and-play training-time framework for On-the-fly Category Discovery (OCD), the setting where a model must label a stream of images from both known and unknown categories using only labeled known-category data. Its central claim is that a pre-trained diffusion model can synthesize images of virtual, unseen categories by interpolating between images of known categories in Stable Diffusion's latent space and CLIP's visual and textual embedding spaces, and that these samples, after filtering and pseudo-labeling, inject category knowledge that existing OCD methods lack. The paper reports that adding DiffGRE to the SMILE baseline gives average gains of 3.7% ACC-ALL and 7.6% ACC-OLD across six fine-grained datasets, and that the combination reaches state-of-the-art accuracy under both hash-like and online clustering inference. If the claim holds, generative pre-training becomes a practical source of training signal for discovering categories no labeled example ever showed.

What carries the argument

The load-bearing operation is spherical interpolation in paired latent spaces, $z^* = \frac{\sin((1-\lambda^*)\theta)}{\sin\theta} z^*_1 + \frac{\sin(\lambda^*\theta)}{\sin\theta} z^*_2$ for $*\in\{t,v,l\}$ with $\theta = \arccos(z^*_1 \cdot z^*_2)$, applied to textual, visual, and diffusion-latent embeddings of two known-category images. This is what turns two known species into a plausible third, unseen species by recombining attributes. Around it sit the diversity filter, which computes cosine similarity between a synthesized image and known-class centers and removes samples above a threshold, and the leader-encoding objective, a contrastive loss that pulls each sample toward its category's mean feature in a high-dimensional space while the same mean features double as the descriptors used for online inference.

What would settle it

A concrete test is to measure, for every retained synthesized image, whether its nearest real-image neighbor belongs to an unknown category; if on another fine-grained dataset most synthesized images are nearest to known-class samples, the ACC-NEW gain over the base OCD method should shrink toward zero. The paper's own ablation already shows the gain depends on this: removing DDR lowers ACC-ALL by 2.3% on average, and a dataset where latent interpolation produces mostly unrealistic blends would be expected to collapse the gain entirely.

Watch

Extended reading notes

Core claim

DiffGRE is a three-stage pipeline added on top of an existing OCD method: Attribute Composition Generation (ACG) performs spherical interpolation between two known-category images in three embedding spaces -- the Stable Diffusion latent space, the CLIP visual space, and the CLIP textual space -- to produce images whose attributes recombine into virtual categories; Diversity-Driven Refinement (DDR) discards synthesized images whose mean cosine similarity to known-class centers exceeds a threshold, keeping only samples that look distinct from known categories; Semi-supervised Leader Encoding (SLE) clusters the union of labeled and synthesized data, aligns the clusters with ground-truth labels via Hungarian assignment, and trains the model with leader-based contrastive learning while maintaining category-level leader features. At inference, SLE's leaders support an online clustering strategy that operates on high-dimensional features rather than binary hash codes. The paper's central claim is that this generate-refine-encode cycle gives OCD models access to category knowledge beyond the labeled set and consistently improves discovery of both known and unknown fine-grained categories.

Load-bearing premise

The claim stands on the assumption that interpolating between two known-category images in the diffusion and CLIP latent spaces produces images that look like plausible members of real unseen categories, so that training on them transfers to the OCD query set.

Editorial extensions

If this is right

  • Adding DiffGRE to SMILE yields average gains of 3.7% ACC-ALL and 7.6% ACC-OLD across six fine-grained datasets, with consistent gains on every dataset under hash-like inference.
  • SMILE + DiffGRE achieves a new state-of-the-art average accuracy under online clustering inference, while PHE + DiffGRE reaches the best average under hash-like inference.
  • SLE-based online cluster inference outperforms hash-descriptor inference: BaseHash + SLE-based beats BaseHash + SLC by an average of 10.8% ACC-ALL and 16.5% ACC-NEW across six datasets.
  • The upper-bound experiment indicates that synthesized virtual categories contribute almost as much as real query data on unknown categories, narrowing the gap to the upper bound from 9.9% to 4.2% ACC-ALL on Mollusca.
  • DiffGRE is plug-and-play: it improves BaseHash, SMILE, and PHE without changing their backbones or requiring access to the query set during training.

Reading between the lines

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

  • If the attribute-composition premise holds broadly, the same generate-refine-encode recipe should transfer to generalized category discovery with offline clustering, where missing-category knowledge is equally the bottleneck; the paper does not test this setting.
  • Because Stable Diffusion's training data may already contain images from the benchmark's unknown categories, part of the observed gain could come from memorization rather than novel attribute composition; measuring performance with a diffusion model whose training set provably excludes the query categories would separate the two.
  • A testable extension is to scale DiffGRE with the scarcity of labels: the upper-bound result suggests its relative gain should grow as the labeled set shrinks, since synthesis supplies the attribute combinations the labeled data lack.
  • The online cluster inference threshold $\Delta_{\max}$ is a single global value derived from maximal intra-category distance; calibrating it per category, or adapting it online as new leaders are created, is a natural refinement the paper leaves implicit.
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 / 6 minor

Summary. DiffGRE is a plug-and-play module for On-the-fly Category Discovery (OCD). It synthesizes virtual-category images by spherical interpolation between two known-class images in the Stable Diffusion latent space and the CLIP text/visual spaces (Attribute Composition Generation, ACG), filters the synthesized images with a Diversity-Driven Refinement (DDR) step that removes samples too close to known class centers, and then pseudo-labels the retained samples through clustering with Hungarian alignment, using a leader-based contrastive loss and an auxiliary classifier during training (Semi-supervised Leader Encoding, SLE). At inference it replaces hash decoding with an Online Cluster Inference (OCI) strategy based on leader features. The paper reports consistent accuracy gains over BaseHash, SMILE, and PHE under both hash-like and OCI inference across six fine-grained datasets, with ablations attributing part of the gain to DDR and SLE.

Significance. If the reported gains are stable and the underlying mechanism is confirmed, the paper makes a practically valuable contribution: it shows how a frozen text-to-image diffusion model can be used as a method-agnostic source of synthetic training data for fine-grained category discovery, without ever accessing the query set during training. Concrete strengths are the breadth of the evaluation (six datasets, three existing OCD methods, two inference protocols), the ablations that isolate ACG, DDR, and SLE, the comparison against MixUp/CutMix/Diff-Mix/Da-Fusion, the computational-cost table, and the publicly released code. The main shortfall is evidential: the paper's central explanatory claim—that the synthetic images carry knowledge about real unseen categories—is supported mainly by two qualitative nearest-neighbor examples and is not measured quantitatively; the paper's own limitation statement and the absence of any error bars leave the scope of the contribution uncertain.

major comments (4)
  1. [§3.1, Eq. (2)] The generation core is not reproducible as written. Eq. (1) produces three interpolated embeddings z_t, z_v, and z_l, and the text says that CLIP visual embeddings are used to 'rectify the noise', but Eq. (2) only displays a denoising objective conditioned on the latent and the text input; there is no equation or pseudocode specifying how z_v is injected into the UNet, at which resolution or layer, and how the interpolated text embedding is combined with the latent during the reverse process. Because ACG is the first and load-bearing component of DiffGRE, please specify the exact conditioning mechanism or provide code-level pseudocode for the denoising loop.
  2. [§4.1, Tables 1–3] All headline numbers are single-run point estimates. The reported improvements (for example, +3.7 ACC-ALL and +7.6 ACC-OLD for SMILE versus SMILE+DiffGRE, and ablation gaps of 2.3–4.6 points in Table 3) are large enough to matter, but no error bars or number of seeds are reported anywhere, so a reader cannot tell whether the gains are stable under the stochasticity of diffusion sampling, clustering, and SGD. Please run at least three seeds for the main comparisons and ablations, report mean and standard deviation, and state whether each conclusion is consistent across seeds.
  3. [§3.1, §3.2, §4.5, Fig. 5] The central explanatory claim—that DiffGRE improves OCD by injecting knowledge about unseen categories—is not yet validated quantitatively. The evidence is a small number of nearest-neighbor examples computed with the full query set, which is unavailable under the OCD training protocol, and no fraction is reported of retained synthetic images that are plausible novel-category members or that match the actual unknown classes. DDR's filter (Eq. 5) keeps everything far from known class centers, which can include undefined objects and incorrect syntheses; the paper's own failure examples in Fig. 5 and the statement in Appendix G that the model has 'unsatisfactory generative results' acknowledge this possibility. Since an alternative explanation is that SLE and the auxiliary classifier act as generic regularizers, please add a diagnostic that distinguishes these interpretations, for example by training the SLE/L_c pipeline on equal numbers of corrupted or randomly mixed images, or by quantitatively measuring pseudo-label precision and category alignment of the retained synthetic samples.
  4. [§4.4, Fig. 4, Appendix B.2] The hyperparameters γ, α, β, λ_v, and λ_l appear to be selected using the target accuracy of the test/query set. Fig. 4(c-d) tunes γ on Arachnida by observing ACC-OLD and ACC-NEW, Table B.2 then fixes γ per dataset by matching the number of retained images to the labeled-set size, and λ_v and λ_l are tuned on CUB. There is no described validation split or protocol that respects the OCD constraint that the query set is unseen during training. Please specify the validation procedure, or state explicitly that no test-set information was used, and report the sensitivity of the main conclusions to γ and the interpolation parameters; if per-dataset tuning is necessary, it should be part of the reported method rather than an after-the-fact model selection.
minor comments (6)
  1. [Table 3] Use consistent decimal notation: '53,2' should be '53.2'.
  2. [Fig. 2 and Sec. 1] The caption and the Introduction say the interpolation is between a 'Laysan albatross' image and another 'Laysan albatross' image; presumably the second should be 'Black-footed albatross'. Please correct.
  3. [§3.4 and Algorithm 1] The threshold Δ_max is never defined; specify how it is computed, and clarify whether distances in Algorithm 1 are L2 or cosine, since Eq. (3) uses cosine similarity.
  4. [§3.3] The 'off-the-shelf clustering method [37]' is not named; specify the clustering algorithm and its hyperparameters so that the SLE pseudo-labeling step is reproducible.
  5. [§3.1, Eq. (2)] The symbol t is overloaded: it denotes both the diffusion timestep and the text prompt in the notation around Eq. (2); rename the text input to avoid ambiguity.
  6. [Appendix B.1, Table A.2] The 'Upbound' experiment trains SLE using the actual query set D_Q as generated data; label it more clearly as an oracle/upper-bound diagnostic and state that this protocol is not an allowed OCD training signal, so that readers do not interpret it as an allowed use of the query set.

Circularity Check

1 steps flagged · score 2.0 of 10

No construction-level circularity; only a mild fitted-hyperparameter effect on the CUB portion of the reported averages.

  1. fitted input called prediction [Sec. 4.4 (Hyper-Parameters Analysis) and Tables 1-2; also App. E]
    "We fine-tuned this parameter on the CUB dataset and fixed it at 0.3 for all datasets. Similarly, β remains stable across variations, as shown in Fig. 4 (b). We set β = 1.0 for all datasets during training. Then, we fix λt to tune the optimal λv = 0.7 and λl = 0.8 (see detailed analysis in the Appendix)."

    The paper reports average gains of 3.7% ACC-ALL and 7.6% ACC-OLD for SMILE + DiffGRE across six datasets, but α, β, and the interpolation weights are selected by evaluating on CUB, and CUB is included in those averages. The CUB contribution is therefore a fitted value, not an independent prediction under a pre-specified configuration. The gains on the remaining datasets are still genuine external measurements (all positive under the transferred settings), so the issue is a mild selection-on-evaluation effect rather than a derivation that reduces to its inputs.

full rationale

The core derivation chain of DiffGRE is otherwise self-contained. ACG (Eq. 1-2) produces synthetic images by spherical interpolation in Stable Diffusion and CLIP spaces; DDR (Eq. 3-5) filters them by cosine distance to known-class centers; SLE (Eq. 6-7) pseudo-labels the retained images by clustering D_S ∪ D_G and Hungarian-aligning to known labels. None of these equations encodes the held-out D_Q labels, and the reported ACC-ALL/OLD/NEW are measured on D_Q with Hungarian alignment (Sec. 4.1). Thus the central empirical claim is an external benchmark measurement, not an equation-level self-definition. Hyperparameter tuning on CUB is a real methodological caveat, but it is confined to a few scalars and the same configuration transfers to five other datasets with consistent positive gains, so it does not rise to a construction-level circularity. No load-bearing self-citation was found: PHE [59] is used as a comparison baseline, and the motivation citations for semantic interpolation are external to this paper's claims.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

DiffGRE introduces no physical entities. The 'virtual categories' are pseudo-labels assigned by clustering, not new real-world classes. The main auxiliary assumptions are about the semantic behavior of pre-trained generative models and the reliability of clustering-based pseudo-labels. The free parameters are standard hyperparameters, but the per-dataset tuning of the DDR threshold is the most consequential choice.

free parameters (4)
  • DDR threshold gamma = 0.20 (Animalia), 0.25 (Pets), 0.30 (Mollusca), 0.40 (CUB, Arachnida), 0.65 (Scars)
    Selected per dataset so that the number of retained synthetic images matches the labeled set size (Appendix B.4, Table B.2). This data-dependent choice directly controls how much synthetic data enters training and varies across datasets.
  • Interpolation parameters lambda_v, lambda_l = 0.7, 0.8
    Empirically tuned on CUB and Arachnida (Appendix E, Fig E.1). lambda_t is fixed at 0.7 following Ramesh et al. The paper states these are selected to balance diversity and ambiguity.
  • Loss weights alpha, beta = 0.3, 1.0
    Alpha is tuned on CUB and fixed for all datasets; beta is tuned on CUB and Stanford Cars (Sec 4.4, Fig 4, Appendix Table B.7). These weights balance the leader-based contrastive loss and the cross-entropy classification loss.
  • Temperature tau = 0.05
    Set empirically in Sec 3.3 and used for all experiments. No sensitivity analysis is provided.
assumptions (5)
  • domain assumption Pre-trained diffusion models have a semantic latent space in which spherical interpolation between images yields meaningful attribute re-compositions.
    Invoked in Sec 1 and 3.1 to justify ACG. Supported only by qualitative examples; no quantitative measure of semantic coherence is provided.
  • ad hoc to paper Composing attributes from seen categories can approximate attributes of unseen categories.
    The core premise of virtual category synthesis. Cited to [24,26,47] but not validated for the OCD benchmarks; the nearest-neighbor success is anecdotal.
  • domain assumption Clustering D_A and aligning clusters to labeled ground truth yields reliable pseudo-labels for generated data.
    Used in Sec 3.3. The paper does not report pseudo-label accuracy or analyze failure modes of the clustering step.
  • domain assumption DINO-pretrained ViT-B-16 features are a suitable backbone for fine-grained OCD.
    Standard choice from SMILE; no comparison with other backbones is given.
  • domain assumption The maximal intra-category distance Delta_max computed from known-class leaders is a valid threshold to detect unknown classes at inference.
    Used in Algorithm 1. No sensitivity analysis for Delta_max is provided; the method assumes this training-derived statistic transfers to the query stream.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generate, Refine, and Encode: Leveraging Synthesized Novel Samples for On-the-Fly Fine-Grained Category Discovery." pith.science (2026). https://pith.science/paper/GJWXRYDO

@misc{pith2026250704051,
  author       = {Pith},
  title        = {Pith review of: Generate, Refine, and Encode: Leveraging Synthesized Novel Samples for On-the-Fly Fine-Grained Category Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GJWXRYDO}},
  note         = {Machine review of arXiv:2507.04051}
}
read the original abstract

In this paper, we investigate a practical yet challenging task: On-the-fly Category Discovery (OCD). This task focuses on the online identification of newly arriving stream data that may belong to both known and unknown categories, utilizing the category knowledge from only labeled data. Existing OCD methods are devoted to fully mining transferable knowledge from only labeled data. However, the transferability learned by these methods is limited because the knowledge contained in known categories is often insufficient, especially when few annotated data/categories are available in fine-grained recognition. To mitigate this limitation, we propose a diffusion-based OCD framework, dubbed DiffGRE, which integrates Generation, Refinement, and Encoding in a multi-stage fashion. Specifically, we first design an attribute-composition generation method based on cross-image interpolation in the diffusion latent space to synthesize novel samples. Then, we propose a diversity-driven refinement approach to select the synthesized images that differ from known categories for subsequent OCD model training. Finally, we leverage a semi-supervised leader encoding to inject additional category knowledge contained in synthesized data into the OCD models, which can benefit the discovery of both known and unknown categories during the on-the-fly inference process. Extensive experiments demonstrate the superiority of our DiffGRE over previous methods on six fine-grained datasets.

Figures

Figures reproduced from arXiv: 2507.04051 by the authors.

Figure 1
Figure 1. Comparison with GCD and OCD. Most existing GCD [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Motivation of this paper. In this case, we randomly [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. (a) Overview of our DiffGRE framework. First, we sample images from different categories to generate virtual-category images [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Impact of hyperparameters, i.e., α, β and γ. We implement experiments on CUB [46] in (a-b) and Arachnida [44] in (c-d). Our Synthesis (Virtual Category) Yellowthroat (Unknown Category) Kentucky Warbler (Known Category) Myrtle Warbler (Known Category) Our Synthesis (Vir…
Figure 5
Figure 5. Figure 5: Visualization of the synthesized images. We illustrate two examples of successful attribute composition on the CUB [ [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 50 canonical work pages

  1. [1]

    Generalized category discovery with decoupled prototypical network

    Wenbin An, Feng Tian, Qinghua Zheng, Wei Ding, QianY- ing Wang, and Ping Chen. Generalized category discovery with decoupled prototypical network. InAAAI, 2023. 3

    work page 2023

  2. [2]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In ICCV, 2021. 6, 2

    work page 2021

  3. [3]

    Parametric information maxi- mization for generalized category discovery

    Florent Chiaroni, Jose Dolz, Ziko Imtiaz Masud, Amar Mitiche, and Ismail Ben Ayed. Parametric information maxi- mization for generalized category discovery. InICCV, 2023. 3

    work page 2023

  4. [4]

    Text-to-image diffusion mod- els are zero shot classifiers.NeurIPS, 2024

    Kevin Clark and Priyank Jaini. Text-to-image diffusion mod- els are zero shot classifiers.NeurIPS, 2024. 3

    work page 2024

  5. [5]

    On-the-fly cate- gory discovery

    Ruoyi Du, Dongliang Chang, Kongming Liang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. On-the-fly cate- gory discovery. InCVPR, 2023. 1, 2, 3, 5, 6, 7, 8, 4

    work page 2023

  6. [6]

    Diffusion-based probabilistic un- certainty estimation for active domain adaptation.NeurIPS,

    Zhekai Du and Jingjing Li. Diffusion-based probabilistic un- certainty estimation for active domain adaptation.NeurIPS,

    work page

  7. [7]

    Diverse data augmentation with diffusions for effective test-time prompt tuning

    Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. InICCV, 2023. 2

    work page 2023

  8. [8]

    A unified objective for novel class discovery

    Enrico Fini, Enver Sangineto, St ´ephane Lathuili `ere, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. InICCV, 2021. 3

    work page 2021

Show all 62 references
  1. [9]

    Dreamda: Generative data augmentation with diffusion mod- els.arXiv preprint arXiv:2403.12803, 2024

    Yunxiang Fu, Chaoqi Chen, Yu Qiao, and Yizhou Yu. Dreamda: Generative data augmentation with diffusion mod- els.arXiv preprint arXiv:2403.12803, 2024. 3

    2024 arXiv

  2. [10]

    An image is worth one word: Personalizing text-to- image generation using textual inversion.arXiv preprint arXiv:2208.01618, 2022

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patash- nik, Amit H Bermano, Gal Chechik, and Daniel Cohen- Or. An image is worth one word: Personalizing text-to- image generation using textual inversion.arXiv preprint arXiv:2208.01618, 2022. 3

    2022 arXiv

  3. [11]

    Learning to discover novel visual categories via deep transfer cluster- ing

    Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer cluster- ing. InICCV, 2019. 3

    2019

  4. [12]

    Autonovel: Automati- cally discovering and learning novel visual categories.IEEE TPAMI, 2021

    Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, An- drea Vedaldi, and Andrew Zisserman. Autonovel: Automati- cally discovering and learning novel visual categories.IEEE TPAMI, 2021. 1, 2, 6

    2021

  5. [13]

    John Wiley & Sons, Inc., 1975

    John A Hartigan.Clustering algorithms. John Wiley & Sons, Inc., 1975. 7, 1

    1975

  6. [14]

    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

    2022 arXiv

  7. [15]

    Denoising diffu- sion probabilistic models.NeurIPS, 2020

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

    2020

  8. [16]

    Diffusemix: Label- preserving data augmentation with diffusion models

    Khawar Islam, Muhammad Zaigham Zaheer, Arif Mah- mood, and Karthik Nandakumar. Diffusemix: Label- preserving data augmentation with diffusion models. In CVPR, 2024. 2, 3

    2024

  9. [17]

    Joint representation learning and novel category discovery on single-and multi-modal data

    Xuhui Jia, Kai Han, Yukun Zhu, and Bradley Green. Joint representation learning and novel category discovery on single-and multi-modal data. InICCV, 2021. 1, 2, 6

    2021

  10. [18]

    Podia- 3d: Domain adaptation of 3d generative model across large domain gap using pose-preserved text-to-image diffusion

    Gwanghyun Kim, Ji Ha Jang, and Se Young Chun. Podia- 3d: Domain adaptation of 3d generative model across large domain gap using pose-preserved text-to-image diffusion. In ICCV, 2023. 3

    2023

  11. [19]

    Proxy anchor-based unsu- pervised learning for continuous generalized category dis- covery

    Hyungmin Kim, Sungho Suh, Daehwan Kim, Daun Jeong, Hansang Cho, and Junmo Kim. Proxy anchor-based unsu- pervised learning for continuous generalized category dis- covery. InICCV, 2023. 3

    2023

  12. [20]

    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 Workshop, 2013. 6

    2013

  13. [21]

    The hungarian method for the assignment problem.Naval research logistics quarterly, 1955

    Harold W Kuhn. The hungarian method for the assignment problem.Naval research logistics quarterly, 1955. 5

    1955

  14. [22]

    Diffusion models already have A semantic latent space

    Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have A semantic latent space. InICLR, 2023. 2

    2023

  15. [23]

    Your diffusion model is secretly a zero-shot classifier

    Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. InICCV, 2023. 3

    2023

  16. [24]

    Context-based and diversity-driven specificity in compositional zero-shot learning

    Yun Li, Zhe Liu, Hang Chen, and Lina Yao. Context-based and diversity-driven specificity in compositional zero-shot learning. InCVPR, 2024. 2

    2024

  17. [25]

    Novel class discovery for ultra-fine-grained visual categorization

    Yu Liu, Yaqi Cai, Qi Jia, Binglin Qiu, Weimin Wang, and Nan Pu. Novel class discovery for ultra-fine-grained visual categorization. InCVPR, 2024. 1

    2024

  18. [26]

    Pmgnet: Disentanglement and entanglement benefit mutually for compositional zero-shot learning.Computer Vision and Image Understanding, 249: 104197, 2024

    Yu Liu, Jianghao Li, Yanyi Zhang, Qi Jia, Weimin Wang, Nan Pu, and Nicu Sebe. Pmgnet: Disentanglement and entanglement benefit mutually for compositional zero-shot learning.Computer Vision and Image Understanding, 249: 104197, 2024. 2

    2024

  19. [27]

    Some methods for classification and analysis of multivariate observations

    James MacQueen et al. Some methods for classification and analysis of multivariate observations. InProceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, 1967. 3

    1967

  20. [28]

    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. 4

    2021 arXiv

  21. [29]

    Cats and dogs

    Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. InCVPR, 2012. 6, 8

    2012

  22. [30]

    Unsupervised domain adap- tation via domain-adaptive diffusion.IEEE TIP, 2024

    Duo Peng, Qiuhong Ke, ArulMurugan Ambikapathi, Yasin Yazici, Yinjie Lei, and Jun Liu. Unsupervised domain adap- tation via domain-adaptive diffusion.IEEE TIP, 2024. 3

    2024

  23. [31]

    Dynamic conceptional contrastive learning for generalized category discovery

    Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptional contrastive learning for generalized category discovery. In CVPR, 2023. 1, 3

    2023

  24. [32]

    Federated generalized category discovery

    Nan Pu, Wenjing Li, Xingyuan Ji, Yalan Qin, Nicu Sebe, and Zhun Zhong. Federated generalized category discovery. In CVPR, 2024. 1

    2024

  25. [33]

    On the momentum term in gradient descent learning algorithms.Neural networks, 1999

    Ning Qian. On the momentum term in gradient descent learning algorithms.Neural networks, 1999. 2

    1999

  26. [34]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InICML, 2021. 2, 3

    2021

  27. [35]

    Hierarchical text-conditional image gener- ation with clip latents.arXiv preprint arXiv:2204.06125, 1 (2):3, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents.arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 8

    2022 arXiv

  28. [36]

    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. InCVPR, 2022. 2, 3, 4

    2022

  29. [37]

    The map equation.The European Physical Journal Special Top- ics, 2009

    Martin Rosvall, Daniel Axelsson, and Carl T Bergstrom. The map equation.The European Physical Journal Special Top- ics, 2009. 5

    2009

  30. [38]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InCVPR, 2023. 3

    2023

  31. [39]

    Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion

    Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion. InCVPR, 2023. 3

    2023

  32. [40]

    Animating rotation with quaternion curves

    Ken Shoemake. Animating rotation with quaternion curves. InProceedings of the 12th annual conference on Computer graphics and interactive techniques, pages 245–254, 1985. 4

    1985

  33. [41]

    Latent space disentanglement in diffusion transformers enables zero-shot fine-grained semantic edit- ing.arXiv preprint arXiv:2408.13335, 2024

    Zitao Shuai, Chenwei Wu, Zhengxu Tang, Bowen Song, and Liyue Shen. Latent space disentanglement in diffusion transformers enables zero-shot fine-grained semantic edit- ing.arXiv preprint arXiv:2408.13335, 2024. 2

    2024 arXiv

  34. [42]

    Effective data augmentation with diffusion models.arXiv preprint arXiv:2302.07944, 2023

    Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models.arXiv preprint arXiv:2302.07944, 2023. 2, 3, 8

    2023 arXiv

  35. [43]

    Test-time stain adaptation with diffusion models for histopathology image classification

    Cheng-Chang Tsai, Yuan-Chih Chen, and Chun-Shien Lu. Test-time stain adaptation with diffusion models for histopathology image classification. InECCV, 2025. 3

    2025

  36. [44]

    The inaturalist species classification and de- tection dataset

    Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and de- tection dataset. InCVPR, 2018. 6, 8, 1

    2018

  37. [45]

    Generalized category discovery

    Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisser- man. Generalized category discovery. InCVPR, 2022. 1, 3

    2022

  38. [46]

    The caltech-ucsd birds-200-2011 dataset.Computation & Neural Systems Technical Report,

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Per- ona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset.Computation & Neural Systems Technical Report,

    2011

  39. [47]

    Videocomposer: Compositional video synthesis with motion controllability.NeurIPS, 2023

    Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Ji- uniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jin- gren Zhou. Videocomposer: Compositional video synthesis with motion controllability.NeurIPS, 2023. 2

    2023

  40. [48]

    Enhance im- age classification via inter-class image mixup with diffusion model

    Zhicai Wang, Longhui Wei, Tan Wang, Heyu Chen, Yanbin Hao, Xiang Wang, Xiangnan He, and Qi Tian. Enhance im- age classification via inter-class image mixup with diffusion model. InCVPR, 2024. 2, 3, 8, 4, 5, 6

    2024

  41. [49]

    Parametric classification for generalized category discovery: A baseline study

    Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. InICCV, 2023. 3

    2023

  42. [50]

    Metagcd: Learning to continually learn in generalized cat- egory discovery

    Yanan Wu, Zhixiang Chi, Yang Wang, and Songhe Feng. Metagcd: Learning to continually learn in generalized cat- egory discovery. InICCV, 2023. 3

    2023

  43. [51]

    Learning to distin- guish samples for generalized category discovery

    Fengxiang Yang, Nan Pu, Li Wenjing, Luo Zhiming, Li Shaozi, Niculae Sebe, and Zhun Zhong. Learning to distin- guish samples for generalized category discovery. InECCV,

  44. [52]

    Diffusion models and semi-supervised learners benefit mutually with few labels.NeurIPS, 2024

    Zebin You, Yong Zhong, Fan Bao, Jiacheng Sun, Chongx- uan Li, and Jun Zhu. Diffusion models and semi-supervised learners benefit mutually with few labels.NeurIPS, 2024. 3

    2024

  45. [53]

    Coca: Contrastive captioners are image-text foundation models.arXiv preprint arXiv:2205.01917, 2022

    Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mo- jtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models.arXiv preprint arXiv:2205.01917, 2022. 4

    2022 arXiv

  46. [54]

    Cutmix: Regu- larization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regu- larization strategy to train strong classifiers with localizable features. InICCV, 2019. 8, 4

    2019

  47. [55]

    mixup: Beyond 903 empirical risk minimization

    Hongyi Zhang, YND Moustapha Cisse, and David Lopez- Paz. mixup: Beyond 903 empirical risk minimization. In ICLR, 2018. 8, 4

    2018

  48. [56]

    Incremental general- ized category discovery

    Bingchen Zhao and Oisin Mac Aodha. Incremental general- ized category discovery. InICCV, 2023. 3

    2023

  49. [57]

    Pyramidal person re-identification via multi-loss dynamic training

    Feng Zheng, Cheng Deng, Xing Sun, Xinyang Jiang, Xi- aowei Guo, Zongqiao Yu, Feiyue Huang, and Rongrong Ji. Pyramidal person re-identification via multi-loss dynamic training. InCVPR, 2019. 5, 2

    2019

  50. [58]

    Textual knowledge matters: Cross-modality co- teaching for generalized visual class discovery

    Haiyang Zheng, Nan Pu, Wenjing Li, Nicu Sebe, and Zhun Zhong. Textual knowledge matters: Cross-modality co- teaching for generalized visual class discovery. InECCV,

  51. [59]

    Prototypical hash encoding for on-the-fly fine- grained category discovery.NeurIPS, 2025

    Haiyang Zheng, Nan Pu, Wenjing Li, Nicu Sebe, and Zhun Zhong. Prototypical hash encoding for on-the-fly fine- grained category discovery.NeurIPS, 2025. 5, 7, 1, 2

    2025

  52. [60]

    Neighborhood contrastive learn- ing for novel class discovery

    Zhun Zhong, Enrico Fini, Subhankar Roy, Zhiming Luo, Elisa Ricci, and Nicu Sebe. Neighborhood contrastive learn- ing for novel class discovery. InCVPR, 2021. 3

    2021

  53. [61]

    Novel class discovery in chest x-rays via paired images and text

    Jiaying Zhou, Yang Liu, and Qingchao Chen. Novel class discovery in chest x-rays via paired images and text. InAAAI,

  54. [62]

    Up- bound

    Yongchao Zhou, Hshmat Sahak, and Jimmy Ba. Training on thin air: Improve image classification with generated data. arXiv preprint arXiv:2305.15316, 2023. 3 Generate, Refine, and Encode: Leveraging Synthesized Novel Samples for On-the-Fly Fine-Grained Category Discovery Supplem...

    2023 arXiv

Pith tools

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