Pith. sign in

REVIEW 5 major objections 8 minor 61 references

TaxaDiffusion: Progressively Trained Diffusion Model for Fine-Grained Species Generation

T0 review · 5 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TaxaDiffusion claims that training a diffusion model progressively from Kingdom to Species yields more faithful, identity-accurate animal images than one-shot conditioning or full fine-tuning, even for species with a single training image.

desk verdict Genuinely new taxonomy-progressive training idea with striking qualitative gains, but the BioCLIP metric is circular and the key ablation does not favor progressive training; deserves refereeing with major revisions. read the letter →

arxiv 2506.01923 v2 pith:BRUCYBN3 submitted 2025-06-02 cs.CV cs.AIcs.LG

Amin Karimi Monsefi , Mridul Khurana , Rajiv Ramnath , Anuj Karpatne , Wei-Lun Chao , Cheng Zhang This is my paper
classification cs.CVcs.AIcs.LG
keywords taxonomy-guideddiffusionprogressivetrainingfine-grainedimagegenerationfew-shotspecieshierarchicalconditioningLinnaeantaxonomymodelsBioCLIPevaluation
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 sets out to show that a diffusion model can generate images of individual animal species, including rare ones with only a handful of training photos, if it is taught the Linnaean taxonomic tree level by level from Kingdom down to Species rather than being asked to learn all species at once. The proposed method, TaxaDiffusion, first adapts a pretrained text-to-image model to a biological dataset, then trains a small conditioning module at each taxonomic rank and freezes it before moving to the next rank. On FishNet, iNaturalist, and BIOSCAN-1M, the paper reports that this progressive schedule outperforms zero-shot Stable Diffusion, LoRA fine-tuning, full fine-tuning, and the prior FineDiffusion approach on image fidelity and taxonomy-image alignment. If the claim is right, biological taxonomy is a sample-efficient inductive bias: shared morphology learned from many species transfers to species with almost no data, which matters for the millions of under-photographed species.

What carries the argument

The load-bearing object is the progressive taxonomy-conditioning cascade: a separate trainable two-layer transformer module for each taxonomic rank takes the frozen CLIP text embedding of the taxonomic name truncated at that rank, and the output embeddings of all trained ranks are summed and fed into the denoising network. A module is frozen as soon as its rank is trained, so later ranks refine rather than overwrite coarser knowledge. The load-bearing inference identity is the modified classifier-free guidance $$\tilde{\epsilon}_\$\theta$(x_t, t, $c^{{(i)}}$) = (1+w)\,\epsilon_\$\theta$(x_t, t, $c^{{(i)}}$) - w\,\epsilon_\$\theta$(x_t, t, $c^{{(0)}}$),$$ which replaces the usual unconditional score with the Kingdom-level score $c^{(0)}$, so generation is steered from a broad biological prior toward the target species in the same order in which training happened.

What would settle it

Have expert taxonomists, or an independently trained classifier, judge whether the generated images of rare species such as Amphichaetodon Melbae, trained from a single photo, match the real species; if they perform at chance while BioCLIP scores stay high, the reported identity-accuracy gains are a metric artifact.

Watch

Extended reading notes

Core claim

TaxaDiffusion's central claim is that the order in which conditioning information is introduced matters as much as the information itself. Starting from Stable Diffusion adapted to biological imagery with LoRA, the method trains a separate two-layer transformer condition module for each of the seven taxonomic ranks, feeding it the frozen CLIP text embedding of the taxonomic name truncated at that rank; once a rank is trained, its module is frozen and the summed embeddings of all trained ranks are passed to the denoising network. Because samples are grouped by their shared ancestor at each stage, the model first learns common body plans, poses, and layouts, and only later the subtle shape, color, and pattern differences that separate families, genera, and species. At inference, the paper modifies classifier-free guidance by subtracting the Kingdom-level score estimate rather than an unconditional estimate, steering generation from a broad biological prior toward the target species. The paper reports that this recipe yields the best FID, LPIPS, and BioCLIP scores on FishNet, iNaturalist, and BIOSCAN-1M, and that species with one to five training images, such as Amphichaetodon Melbae with a single sample, can be generated with recognizable traits through transfer from sibling species.

Load-bearing premise

The load-bearing premise is that BioCLIP scores measure true species-specific morphology, but BioCLIP was trained on the same kind of taxonomy-labeled naturalist imagery used for evaluation, and the paper never tests whether high scores could come from generic taxonomy-consistent images.

Editorial extensions

If this is right

  • Species with only one to a handful of training images become generable, because coarser ranks supply shared body plans and only the final rank must be specialized.
  • The model can generate images at any taxonomic rank, so shared family- or genus-level traits can be visualized without sampling all descendant species.
  • Freezing each rank's condition module before training the next prevents later stages from erasing the shared-structure knowledge learned earlier.
  • Replacing the unconditional score in classifier-free guidance with the Kingdom-level score improves both image fidelity and taxonomy-image alignment.
  • These gains come while training only about 6.5% of the model's parameters, far fewer than full fine-tuning requires.

Reading between the lines

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

  • A natural extension, not tested in the paper, is to apply the same progressive hierarchy to any domain with a known class tree, such as plant varieties, minerals, or manufactured parts, where coarse classes share structure and fine classes differ subtly.
  • Because BioCLIP and the evaluation images share the same taxonomy-labeled naturalist distribution, an independent trait-level test, such as checking for the presence or absence of a known diagnostic stripe, would give stronger evidence about whether generated images encode real species morphology.
  • A controlled synthetic-taxonomy experiment, with visual features engineered to match the hierarchy, could separate the benefit of the progressive schedule from the benefit of the taxonomy text embeddings themselves.
  • The paper names transformer backbones as future work; combining progressive taxonomy conditioning with a diffusion transformer is a direct scalability path that could improve rare-species generation further.
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

5 major / 8 minor

Summary. The paper proposes TaxaDiffusion, a training framework for fine-grained species image generation. Building on Stable Diffusion v1.5 and LoRA-based domain adaptation, it adds per-taxonomic-level conditioning modules (a frozen CLIP text encoder followed by two transformer layers) and trains them progressively from Kingdom down to Species, with the standard classifier-free guidance term replaced by guidance toward the Kingdom-level condition. The authors claim that this hierarchical progressive training yields lower FID, lower LPIPS, and higher BioCLIP alignment than vanilla Stable Diffusion, SD+LoRA, SD+Full, and FineDiffusion on FishNet, iNaturalist, and BIOSCAN-1M, and that it enables accurate generation of rare species with very few training samples. The paper includes ablations of the training strategy (All vs Random vs Progressive), of the taxonomically guided inference, and qualitative analyses including attention maps, t-SNE embeddings, and failure cases in Appendix C.

Significance. If the central claim holds, the paper makes a useful contribution: it offers a domain-knowledge-informed training curriculum that is parameter-efficient (about 6.5% of the model trainable, per Table S1), spans three taxonomically diverse datasets, and includes several thoughtful analyses (guidance-scale sweep, attention visualizations at different taxonomic levels, and an honest failure-case appendix). I give explicit credit for the clear formulation of the progressive-training idea, the multi-dataset scope, and the documented Dipneusti failure in Appendix C, which is the kind of reporting that makes a paper credible. However, the experimental support for the headline claims is not yet strong enough: the decisive progressive-vs-random ablation goes the wrong way on FID (Table 4), the only identity metric (BioCLIP) is not independent of the evaluation domain, and no comparison is error-barred or compute-matched. The contribution is potentially valuable but not currently established to the standard that the abstract's claims require.

major comments (5)
  1. [Table 4, §4.3] The progressive-training ablation, which is the decisive evidence for the paper's central novelty, shows that Random-level training achieves a better FID (29.53) than Progressive training (31.87), while the BioCLIP advantage is only 10.43 vs 9.85. The text in §4.3 dismisses the FID deficit with the claim that random training 'fails to match the detail and specificity achieved by TaxaDiffusion', but no metric, figure, or statistical test is provided to support that sentence. Since the abstract, Section 1, and the Remark all state that progressive training is the core contribution, the ablation as reported does not support the superiority claim. The authors should report per-run variance, test whether the FID and BioCLIP differences are significant, add an independent identity or fidelity metric, and show per-species results, or explicitly temper the claim.
  2. [§4.1, Evaluation Metrics; Tables 1-3] BioCLIP is the only identity-alignment metric, and it was trained on TreeOfLife-10M, which is assembled from the same kind of taxonomy-labeled naturalist imagery (including iNaturalist) used for the iNaturalist and FishNet evaluations. The paper reports only mean cosine similarity between generated images and taxonomic text prompts and never checks whether a high score corresponds to the correct species. Species-level BioCLIP scores are low (10.43 for TaxaDiffusion vs 8.31 for SD+Full in Table 1), and the gaps at species level are small. A model that produces generic 'fish-like' images could score moderately at higher levels while failing at the species level; the reported identity gains are therefore ambiguous. The authors should add at least one independent identity test, such as top-1/top-5 species classification accuracy of generated images with a held-out classifier, retrieval-based verification, or a human evaluation.
  3. [§4.2, Tables 1-3 and Table 3 (FineDiffusion)] No table reports standard deviations or significance tests, and the FineDiffusion comparison in Table 3 does not specify which 171 species were used, how many images per species were generated, the size and construction of the FID reference set, or the compute budget (FineDiffusion uses DiT-XL/2 and is trained on iNaturalist; TaxaDiffusion uses SD 1.5). The abstract's claim that TaxaDiffusion 'outperforms existing approaches' requires a protocol-matched and compute-matched comparison. At minimum, the caption of Table 1 should state how FID is computed (which reference images, how many, and whether the 10 generated images per category are the same across baselines), and the FineDiffusion row should report the equalized-evaluation protocol and error bars.
  4. [Abstract, §4.3, Appendix C and Table S2] The abstract and Section 1 claim that TaxaDiffusion enables accurate generation 'even with limited training samples per species', but the only supporting evidence is anecdotal (Figure 2, with species having 1-5 samples). Appendix C explicitly reports that the Dipneusti class (5 species, 19 samples) yields suboptimal generations, and Table S2 shows FID 159.77 and BioCLIP 8.87 for that class, which is direct evidence against the unqualified sample-efficiency claim. The authors should either scope the claim to taxa with a sufficient number of related species or provide a controlled study that varies per-species sample counts while holding everything else fixed.
  5. [Remark (§1) and §3.2] A load-bearing premise of the method is that the Linnaean text hierarchy, embedded by frozen CLIP, captures visual similarity among species. The Remark asserts this ('it does capture visual similarity'), and the training procedure depends on it, but the paper never ablates the hierarchy itself. The t-SNE analysis in Figure S2 is suggestive but is shown only at the Family level and compares TaxaDiffusion embeddings with raw CLIP embeddings, not with an alternative conditioning hierarchy. The authors should test whether the progressive benefit comes from taxonomy per se or from the curriculum order, for example by comparing against training with shuffled or reversed taxonomic levels, or with an arbitrary text hierarchy (e.g., alphabetical genus names).
minor comments (8)
  1. [Abstract] The abstract contains the grammatical error 'demonstrate that outperforms existing approaches' (missing subject), which should be corrected.
  2. [§5, Conclusion] The conclusion introduces 'TaxaGuide' as a method component, but this term is never defined in the paper; the method is called TaxaDiffusion elsewhere. Please define or remove the term.
  3. [Appendix B.1 heading] The heading 'Detailed Results on More Spices' should read 'More Species'.
  4. [§4.3, Trait Discovery] The text refers to order 'Acanthuriformers'; the correct spelling is 'Acanthuriformes'.
  5. [Table S1 and §A.1] The model size for TaxaDiffusion (915.45M) exceeds both vanilla SD (859.52M) and SD+LoRA (860.32M) even though only 55.93M parameters are trainable; please clarify what the added 55-56M non-trainable-or-additional parameters are (presumably the per-level conditioning modules) and why the total is 915.45M rather than 860.32M plus the conditioning modules.
  6. [Tables 1-3, captions] Please state explicitly what the BioCLIP score is: whether the reported numbers are mean cosine similarities scaled by 100, over how many generated images per level, and with which prompt template.
  7. [Footnote 1 (§3.2) vs §A.2] Footnote 1 says the LoRA is trained only at the Kingdom level and then frozen, while §4.1 (Implementation Details) says the learning rate 1e-4 is used 'for LoRA and the first taxonomy layer'. Please clarify whether the LoRA and the Kingdom-level module are trained jointly or sequentially, and whether the LoRA remains trainable during later levels.
  8. [§2, Related Work] Reference [20] (Khurana et al., hierarchical conditioning of diffusion models using tree-of-life) is closely related and is cited in the introduction but not discussed in Related Work; a brief positioning against it would help the reader.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation-based circularity: TaxaDiffusion's claims are empirical, tested against external metrics and controlled ablations, and the self-citations are peripheral rather than load-bearing.

full rationale

The paper's central claim is that progressively training a conditioned diffusion model through taxonomic levels improves fine-grained species generation. This claim is supported by quantitative comparisons against external baselines (SD, SD+LoRA, SD+Full, and FineDiffusion) using FID, LPIPS, and BioCLIP, and by ablations comparing All, Random, and Progressive training strategies. No parameter is fitted to the evaluation metric, and no evaluation score is used as a training objective. The inference-time guidance modification in Equation 6 is a design choice, not a derivation that reduces the result to an input. The progressive conditioning modules are trained on taxonomic text labels, while the metrics are computed on generated images against real-image statistics and BioCLIP embeddings. Self-citations to prior work (e.g., [20], [27], [28], [46], [50]) appear only as related work or motivation and are not used to justify the paper's core empirical findings; there is no imported uniqueness theorem or ansatz that the paper treats as externally forced. The concern that BioCLIP was trained on TreeOfLife-10M, which contains similar taxonomy-labeled naturalist imagery to the evaluation datasets, is a legitimate benchmark-validity question, but it is not a circularity of the paper's derivation: the same external metric is applied uniformly to all methods, and the paper does not claim to derive BioCLIP from its own model. Therefore, no specific circular step can be quoted or exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities. 'TaxaGuide' in Section 5 is a label for the guidance formulation in Eq. 6, not a new postulated object. The count of free parameters reflects the main hyperparameters whose values drive the reported results, and the axiom list captures the domain assumptions that must hold for the progressive training story to work.

free parameters (3)
  • guidance scale w = 6
    Chosen by sweeping 0-10 in Appendix B.5; all main results use this value, and the method's quality depends strongly on it.
  • LoRA rank = 4
    Set in all experiments (Appendix A.2); no sensitivity analysis is reported.
  • training iterations per taxonomic level = 250K
    Each progressive level is trained for 250K iterations (Section 4.1), far exceeding the 100K used for the full fine-tuning baseline; this confounds the comparison.
assumptions (5)
  • domain assumption Linnaean taxonomy levels correspond to nested visual similarity that can be exploited for knowledge transfer.
    Section 1 and Figure 1 assert that species sharing ancestors share visual traits; the whole method relies on this mapping.
  • domain assumption Frozen CLIP text embeddings of taxonomic names, after two transformer layers, can represent level-specific visual traits.
    Section 3.2 builds the conditioning modules on the CLIP text encoder; there is no direct evidence that text-only taxonomic names encode morphology.
  • ad hoc to paper Freezing previously trained level modules prevents interference and preserves coarse-to-fine knowledge.
    Section 3.2 states the freezing mechanism stabilizes learning; this design choice is not ablated independently.
  • standard math Standard DDPM denoising equations and Stable Diffusion v1.5 conditioning are valid background.
    Equations 1-3 are taken from prior diffusion literature and are not re-derived.
  • domain assumption BioCLIP score is a valid proxy for taxonomic identity accuracy of generated images.
    Section 4.1 uses BioCLIP as the main alignment metric without discussing its training data overlap with evaluation sets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TaxaDiffusion: Progressively Trained Diffusion Model for Fine-Grained Species Generation." pith.science (2026). https://pith.science/paper/BRUCYBN3

@misc{pith2026250601923,
  author       = {Pith},
  title        = {Pith review of: TaxaDiffusion: Progressively Trained Diffusion Model for Fine-Grained Species Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BRUCYBN3}},
  note         = {Machine review of arXiv:2506.01923}
}
read the original abstract

We propose TaxaDiffusion, a taxonomy-informed training framework for diffusion models to generate fine-grained animal images with high morphological and identity accuracy. Unlike standard approaches that treat each species as an independent category, TaxaDiffusion incorporates domain knowledge that many species exhibit strong visual similarities, with distinctions often residing in subtle variations of shape, pattern, and color. To exploit these relationships, TaxaDiffusion progressively trains conditioned diffusion models across different taxonomic levels -- starting from broad classifications such as Class and Order, refining through Family and Genus, and ultimately distinguishing at the Species level. This hierarchical learning strategy first captures coarse-grained morphological traits shared by species with common ancestors, facilitating knowledge transfer before refining fine-grained differences for species-level distinction. As a result, TaxaDiffusion enables accurate generation even with limited training samples per species. Extensive experiments on three fine-grained animal datasets demonstrate that outperforms existing approaches, achieving superior fidelity in fine-grained animal image generation. Project page: https://amink8.github.io/TaxaDiffusion/

Figures

Figures reproduced from arXiv: 2506.01923 by the authors.

Figure 1
Figure 1. Taxonomy encodes a rich hierarchical structure for cat [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Generative examples of our approach on the FishNet dataset [19]. As we progress through the taxonomy tree from Class to Order, Family, Genus, and finally Species, our model refines its understanding of distinguishing traits, generating realistic im￾ages that capture the unique visual characteristics at each level. For rare species with limited training samples, such as “Amphichaetodon Howensis” (4 samples), “Amphich… view at source ↗
Figure 3
Figure 3. Overview of TaxaDiffusion for progressive diffusion model training. Left and middle: We incorporate a taxonomy-informed conditioning module to guide the model training. Right: The hierarchical taxonomy levels (i.e., Kingdom, Phylum, Class, Order, Family, Genus, Species) are progressively encoded by the text encoder of the CLIP model during training. Each taxonomy level activates specific traits, allowing the model t… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on FishNet [19]. We show four samples from each method and observe that TaxaDiffusion captures fine-grained traits (fin structure, body shape, and pattern) more accurately than the baselines which fail to generate the correct species. fine-graine…
Figure 5
Figure 5. Figure 5: Qualitative results on iNaturalist [48]. We show sam￾ples from TaxaDiffusion generating morphologically correct traits with diverse background and orientations at the Genus level [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Ablation of TaxaDiffusion. Comparison of generated images using different training strategies in taxonomy-driven con￾ditional diffusion models. The columns represent four species [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 32 canonical work pages

  1. [20]

    Hierarchical conditioning of diffusion models using tree-of-life for study- ing species evolution

    Mridul Khurana, Arka Daw, M Maruf, Josef C Uyeda, Wasila Dahdul, Caleb Charpentier, Yasin Bakıs ¸, Henry L Bart Jr, Paula M Mabee, Hilmar Lapp, et al. Hierarchical conditioning of diffusion models using tree-of-life for study- ing species evolution. 2024. 2, 3

    work page 2024

  2. [1]

    idigbio.http://www.idigbio.org/portal, 2020. 17

    work page 2020

  3. [2]

    Inhs collections data.http://biocoll.inhs.illinois.edu/portal/index.php,

    work page

  4. [3]

    Generative novel view synthesis with 3d-aware diffusion models

    Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4217–4229, 2023. 1

    work page 2023

  5. [4]

    Hierarchical integration diffusion model for realistic image deblurring.Advances in neural information processing systems, 36, 2024

    Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring.Advances in neural information processing systems, 36, 2024. 3

    work page 2024

  6. [5]

    Very deep vaes generalize autoregressive mod- els and can outperform them on images.arXiv preprint arXiv:2011.10650, 2020

    Rewon Child. Very deep vaes generalize autoregressive mod- els and can outperform them on images.arXiv preprint arXiv:2011.10650, 2020. 3

    arXiv 2011

  7. [6]

    Diffusion models beat gans on image synthesis.Advances in neural informa- tion processing systems, 34:8780–8794, 2021

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.Advances in neural informa- tion processing systems, 34:8780–8794, 2021. 1, 3

    work page 2021

  8. [7]

    Phylogeny, taxonomy and nomenclature: The problem of taxonomic categories and of nomenclatural ranks.Zootaxa, 1519:27–68, 2007

    Alain Dubois. Phylogeny, taxonomy and nomenclature: The problem of taxonomic categories and of nomenclatural ranks.Zootaxa, 1519:27–68, 2007. 2

    work page 2007

Show all 61 references
  1. [8]

    Make-a-scene: Scene- based text-to-image generation with human priors

    Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene- based text-to-image generation with human priors. InEu- ropean Conference on Computer Vision, pages 89–106. Springer, 2022. 3

    2022

  2. [9]

    A step towards worldwide biodiversity assessment: The bioscan-1m insect dataset.Advances in Neural Information Processing Systems, 36, 2024

    Zahra Gharaee, ZeMing Gong, Nicholas Pellegrino, Iuliia Zarubiieva, Joakim Bruslund Haurum, Scott Lowe, Jaclyn McKeown, Chris Ho, Joschka McLeod, Yi-Yun Wei, et al. A step towards worldwide biodiversity assessment: The bioscan-1m insect dataset.Advances in Neural Information P...

    2024

  3. [10]

    Generative adversarial nets.Advances in neural information processing systems, 27, 2014

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.Advances in neural information processing systems, 27, 2014. 3

    2014

  4. [11]

    Highly personalized text embedding for image manipulation by stable diffusion.arXiv preprint arXiv:2303.08767, 2023

    Inhwa Han, Serin Yang, Taesung Kwon, and Jong Chul Ye. Highly personalized text embedding for image manipulation by stable diffusion.arXiv preprint arXiv:2303.08767, 2023. 3

    2023 arXiv

  5. [12]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3, 4, 8

    2021

  6. [13]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022. 4

    2022 arXiv

  7. [14]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 1, 3

    2020

  8. [15]

    Barik, Les Christidis, Stephen T.Garnett, Paul Kirk, Thomas M

    Donald Hobern, Saroj K. Barik, Les Christidis, Stephen T.Garnett, Paul Kirk, Thomas M. Orrell, Thomas Pape, Richard L. Pyle, Kevin R. Thiele, Frank E. Zachos, and Olaf B´anki. Towards a global list of accepted species vi: The cat- alogue of life checklist.Organisms Diversity &...

    2021

  9. [16]

    Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021. 1, 4, 5, 6

    2021 arXiv

  10. [17]

    Diffusion model-based image editing: A survey.arXiv preprint arXiv:2402.17525, 2024

    Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey.arXiv preprint arXiv:2402.17525, 2024. 3

    2024 arXiv

  11. [18]

    Guiding a dif- fusion model with a bad version of itself.arXiv preprint arXiv:2406.02507, 2024

    Tero Karras, Miika Aittala, Tuomas Kynk ¨a¨anniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a dif- fusion model with a bad version of itself.arXiv preprint arXiv:2406.02507, 2024. 5

    2024 arXiv

  12. [19]

    Fishnet: A large-scale dataset and bench- mark for fish recognition, detection, and functional trait pre- diction

    Faizan Farooq Khan, Xiang Li, Andrew J Temple, and Mo- hamed Elhoseiny. Fishnet: A large-scale dataset and bench- mark for fish recognition, detection, and functional trait pre- diction. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 20496–2...

    2023

  13. [21]

    Diffu- sionclip: Text-guided diffusion models for robust image ma- nipulation

    Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffu- sionclip: Text-guided diffusion models for robust image ma- nipulation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426–2435,

  14. [22]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Diederik P Kingma. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013. 3

    2013 arXiv

  15. [23]

    Controlnet++: Improving conditional controls with efficient consistency feedback

    Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaon- ing Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. InEuropean Conference on Computer Vision, pages 129–147. Springer, 2025. 3

    2025

  16. [24]

    Progressive generation of 3d point clouds with hierarchical consistency.Pattern Recognition, 136:109200, 2023

    Peipei Li, Xiyan Liu, Jizhou Huang, Deguo Xia, Jianzhong Yang, and Zhen Lu. Progressive generation of 3d point clouds with hierarchical consistency.Pattern Recognition, 136:109200, 2023. 3

    2023

  17. [25]

    When stylegan meets stable diffusion: a w+ adapter for person- alized image generation

    Xiaoming Li, Xinyu Hou, and Chen Change Loy. When stylegan meets stable diffusion: a w+ adapter for person- alized image generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2187–2196, 2024. 3

    2024

  18. [26]

    Fine-grained image editing by pixel-wise guidance using diffusion models.arXiv preprint arXiv:2212.02024, 2022

    Naoki Matsunaga, Masato Ishii, Akio Hayakawa, Kenji Suzuki, and Takuya Narihira. Fine-grained image editing by pixel-wise guidance using diffusion models.arXiv preprint arXiv:2212.02024, 2022. 3

    2022 arXiv

  19. [27]

    Detailclip: Detail-oriented clip for fine-grained tasks.arXiv preprint arXiv:2409.06809, 2024

    Amin Karimi Monsefi, Kishore Prakash Sailaja, Ali Alilooee, Ser-Nam Lim, and Rajiv Ramnath. Detailclip: Detail-oriented clip for fine-grained tasks.arXiv preprint arXiv:2409.06809, 2024. 1 9

    2024 arXiv

  20. [28]

    Knobgen: Controlling the sophistication of artwork in sketch-based dif- fusion models.arXiv preprint arXiv:2410.01595, 2024

    Pouyan Navard, Amin Karimi Monsefi, Mengxi Zhou, Wei- Lun Chao, Alper Yilmaz, and Rajiv Ramnath. Knobgen: Controlling the sophistication of artwork in sketch-based dif- fusion models.arXiv preprint arXiv:2410.01595, 2024. 1, 3

    2024 arXiv

  21. [29]

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

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

    2021 arXiv

  22. [30]

    The integrative future of taxonomy.Fron- tiers in zoology, 7:1–14, 2010

    Jos ´e M Padial, Aur ´elien Miralles, Ignacio De la Riva, and Miguel Vences. The integrative future of taxonomy.Fron- tiers in zoology, 7:1–14, 2010. 2

    2010

  23. [31]

    Finediffusion: Scaling up diffusion models for fine- grained image generation with 10,000 classes.arXiv preprint arXiv:2402.18331, 2024

    Ziying Pan, Kun Wang, Gang Li, Feihong He, and Yongxuan Lai. Finediffusion: Scaling up diffusion models for fine- grained image generation with 10,000 classes.arXiv preprint arXiv:2402.18331, 2024. 3, 5, 7, 14

    2024 arXiv

  24. [32]

    Precisecontrol: En- hancing text-to-image diffusion models with fine-grained at- tribute control

    Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Te- jan Karmali, and R Venkatesh Babu. Precisecontrol: En- hancing text-to-image diffusion models with fine-grained at- tribute control. InEuropean Conference on Computer Vision, pages 469–487. Springer, 2025. 3

    2025

  25. [33]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 4195–4205,

  26. [34]

    Hierarchical generation of human-object inter- actions with diffusion probabilistic models

    Huaijin Pi, Sida Peng, Minghui Yang, Xiaowei Zhou, and Hujun Bao. Hierarchical generation of human-object inter- actions with diffusion probabilistic models. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 15061–15073, 2023. 3

    2023

  27. [35]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023. 1

    2023 arXiv

  28. [36]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, p...

    2021

  29. [37]

    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. 3

    2022 arXiv

  30. [38]

    Gener- ating diverse high-fidelity images with vq-vae-2.Advances in neural information processing systems, 32, 2019

    Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Gener- ating diverse high-fidelity images with vq-vae-2.Advances in neural information processing systems, 32, 2019. 3

    2019

  31. [39]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2, 3, 5, 6, 7, 12

    2022

  32. [40]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500–...

    2023

  33. [41]

    Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat-...

    2024

  34. [42]

    Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information ...

    2022

  35. [43]

    Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in Neural In- formation Processing Systems, 35:25278–25294, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, et al. Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in Neural In- fo...

    2022

  36. [44]

    pytorch-fid: FID Score for PyTorch

    Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid,

  37. [45]

    Bioclip: A vision foundation model for the tree of life

    Samuel Stevens, Jiaman Wu, Matthew J Thompson, Eliza- beth G Campolongo, Chan Hee Song, David Edward Carlyn, Li Dong, Wasila M Dahdul, Charles Stewart, Tanya Berger- Wolf, et al. Bioclip: A vision foundation model for the tree of life. InProceedings of the IEEE/CVF Conference ...

  38. [46]

    An em- pirical study and analysis of text-to-image generation us- ing large language model-powered textual representation

    Zhiyu Tan, Mengping Yang, Luozheng Qin, Hao Yang, Ye Qian, Qiang Zhou, Cheng Zhang, and Hao Li. An em- pirical study and analysis of text-to-image generation us- ing large language model-powered textual representation. In European Conference on Computer Vision, pages 472–489. ...

    2024

  39. [47]

    Nvae: A deep hierarchical vari- ational autoencoder.Advances in neural information pro- cessing systems, 33:19667–19679, 2020

    Arash Vahdat and Jan Kautz. Nvae: A deep hierarchical vari- ational autoencoder.Advances in neural information pro- cessing systems, 33:19667–19679, 2020. 3

    2020

  40. [48]

    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. InProceedings of the IEEE conference on computer vision and pattern recognition, pages ...

  41. [49]

    Animatelcm: Accelerating the animation of personalized diffusion mod- els and adapters with decoupled consistency learning.arXiv preprint arXiv:2402.00769, 2024

    Fu-Yun Wang, Zhaoyang Huang, Xiaoyu Shi, Weikang Bian, Guanglu Song, Yu Liu, and Hongsheng Li. Animatelcm: Accelerating the animation of personalized diffusion mod- els and adapters with decoupled consistency learning.arXiv preprint arXiv:2402.00769, 2024. 3

    2024 arXiv

  42. [50]

    To- wards effective usage of human-centric priors in diffusion models for text-based human image generation

    Junyan Wang, Zhenhong Sun, Zhiyu Tan, Xuanbai Chen, Weihua Chen, Hao Li, Cheng Zhang, and Yang Song. To- wards effective usage of human-centric priors in diffusion models for text-based human image generation. InProceed- ings of the IEEE/CVF Conference on Computer Vision and P...

    2024

  43. [51]

    Fg-t2m: Fine-grained text-driven human motion generation via diffusion model

    Yin Wang, Zhiying Leng, Frederick WB Li, Shun-Cheng Wu, and Xiaohui Liang. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 22035–22044, 2023. 3

    2023

  44. [52]

    Generative hierarchical features from synthe- sizing images

    Yinghao Xu, Yujun Shen, Jiapeng Zhu, Ceyuan Yang, and Bolei Zhou. Generative hierarchical features from synthe- sizing images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4432– 4442, 2021. 3

    2021

  45. [53]

    Fine-grained appearance transfer with diffusion mod- els.arXiv preprint arXiv:2311.16513, 2023

    Yuteng Ye, Guanwen Li, Hang Zhou, Cai Jiale, Junqing Yu, Yawei Luo, Zikai Song, Qilong Xing, Youjia Zhang, and Wei Yang. Fine-grained appearance transfer with diffusion mod- els.arXiv preprint arXiv:2311.16513, 2023. 3

    2023 arXiv

  46. [54]

    Chatface: Chat-guided real face edit- ing via diffusion latent space manipulation.arXiv preprint arXiv:2305.14742, 2023

    Dongxu Yue, Qin Guo, Munan Ning, Jiaxi Cui, Yuesheng Zhu, and Li Yuan. Chatface: Chat-guided real face edit- ing via diffusion latent space manipulation.arXiv preprint arXiv:2305.14742, 2023. 3

    2023 arXiv

  47. [55]

    Dilightnet: Fine-grained light- ing control for diffusion-based image generation

    Chong Zeng, Yue Dong, Pieter Peers, Youkang Kong, Hongzhi Wu, and Xin Tong. Dilightnet: Fine-grained light- ing control for diffusion-based image generation. InACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 3

    2024

  48. [56]

    Iti- gen: Inclusive text-to-image generation

    Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Iti- gen: Inclusive text-to-image generation. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 3969–3980, 2023. 3

    2023

  49. [57]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836–3847, 2023. 1, 3

    2023

  50. [58]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 5

    2018

  51. [59]

    A survey on personalized content synthesis with diffusion models.arXiv preprint arXiv:2405.05538, 2024

    Xulu Zhang, Xiao-Yong Wei, Wengyu Zhang, Jinlin Wu, Zhaoxiang Zhang, Zhen Lei, and Qing Li. A survey on personalized content synthesis with diffusion models.arXiv preprint arXiv:2405.05538, 2024. 3

    2024

  52. [60]

    Sine: Single image editing with text- to-image diffusion models

    Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text- to-image diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027–6037, 2023. 1

    2023

  53. [61]

    Actinopteri

    Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models.Advances in Neural Information Processing Sys- tems, 36, 2024. 3 11 T a x a D i f f u s i o n :Progressively ...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.