Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

SliderSpace: Decomposing the Visual Capabilities of Diffusion Models

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read Given a single text prompt, SliderSpace automatically discovers interpretable, composable control directions that decompose concepts, expose art styles, and restore a distilled model's lost diversity.

desk verdict A useful, well-engineered method for unsupervised discovery of composable control directions in diffusion models, but the claimed orthogonality guarantee is not supported by the objective as written and needs measurement or an explicit penalty. read the letter →

arxiv 2502.01639 v1 pith:WRUHWO4A submitted 2025-02-03 cs.CV cs.GRcs.LG

classification cs.CVcs.GRcs.LG
keywords diffusionmodelsunsuperviseddirectiondiscoverylow-rankadaptersCLIPembeddingsprincipalcomponentanalysissemanticdecompositionmodecollapsetext-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

SliderSpace claims that the visual variety a diffusion model can produce for a single text prompt has a discoverable structure: a small set of dominant semantic axes of variation. The method finds those axes with no attribute specified by the user — sample the model, embed the outputs in CLIP space, a semantic visual-embedding space, take principal components, and train one low-rank adapter per component so that turning it on shifts the image in exactly that semantic direction. If the claim holds, anyone can explore and compose the axes of a model's knowledge of a concept by moving sliders, without having to name the attributes in advance. The paper demonstrates this on concepts like 'monster' and 'toy', shows its discovered artistic directions match a manually curated list of 4,388 artist styles by FID, and improves a distilled model's FID from 15.52 to 12.12, nearly matching the undistilled original at 11.72.

What carries the argument

The machinery has three parts working in sequence. Final-image extrapolation (Eq. 2) estimates, from a single noise prediction, the complete image $\tilde{x}_{0,t}$ the model is moving toward at a given denoising step $t$, so that thousands of candidate images can be inspected without running full denoising passes. Semantic decomposition (Eq. 4) runs PCA on the CLIP embeddings of those estimates, yielding orthonormal directions $v_i$ that rank the model's variations by magnitude. Slider training (Eq. 5) attaches a low-rank adapter (LoRA: a lightweight weight update confined to a small-rank subspace) $T_i$ for each direction and optimizes $\mathcal{L} = \sum_i (1 - \cos(\Delta\phi_i, v_i))$, where $\Delta\phi_i$ is the adapter's displacement in CLIP space, forcing each adapter to reproduce exactly one principal component. The PCA step carries the argument: it converts random sampling into a catalogue of nameable axes, and its orthogonality is what keeps the sliders independent.

What would settle it

A decisive control replaces the PCA directions with random orthogonal directions in CLIP space and trains sliders against them: if users find the random-direction sliders just as interpretable and diverse, the spectral-decomposition step is not carrying the method. The paper's own appendix already shows the discovered directions depend heavily on the choice of encoder (CLIP versus FaceNet), so the same experiment can be run with an encoder chosen to be blind to a known axis of variation to test whether that axis vanishes from the sliders.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a diffusion model's response to a text prompt is not an unstructured pile of random variations but a low-dimensional semantic manifold whose principal axes are recoverable and controllable. SliderSpace generates about five thousand images for the prompt, uses final-image extrapolation to read off the image the model is 'planning' at each denoising step, embeds those estimates with CLIP, and takes the principal components of the embedding cloud. Each principal component becomes a slider: a rank-one LoRA adapter on the model's cross-attention layers, trained with a cosine-similarity objective so that attaching the slider moves the output in CLIP space along that component. Because the components are mutually orthogonal, the sliders are semantically orthogonal and composable; because the directions come from the model's own samples rather than user-typed attributes, the decomposition is unsupervised and can reveal axes of variation — sub-concepts, styles, materials, moods — that the model never names in text.

Load-bearing premise

The load-bearing premise, which the paper itself concedes in its limitations section, is that the principal components of CLIP embeddings of the model's own samples point along the axes that humans would call the model's visual capabilities; if that alignment fails, the sliders decompose the encoder's view of the concept rather than the model's.

Editorial extensions

If this is right

  • A user can explore a concept the way they would pull sliders in an editor: turning one direction up, or combining several, yields variants of 'monster' that range from spiky metallic armor to bestial fire and glow, without typing a single attribute.
  • The sliders transfer beyond the prompt they were trained on: directions discovered for 'person' still work on 'police', 'athlete', and even the out-of-domain 'dog', which suggests they capture general visual transformations rather than prompt-specific overfitting.
  • Mode collapse in fast distilled models can be reversed: randomly activating 3 of the 64 sliders on the 4-step SDXL-DMD model improves FID-30k from 15.52 to 12.12, close to the undistilled model's 11.72, while keeping the four-step generation speed.
  • A single decomposition of 'artwork in the style of a famous artist' covers the artistic styles in the model's training data about as well as a manually curated list of 4,388 real artist names, and beats supervised attribute-specified sliders on FID (19.12 versus 32.86).
  • Ten automatically discovered directions already match the FID of 64 attribute-specified concept sliders, so unsupervised discovery packs more variation per parameter than hand-specified controls.

Reading between the lines

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

  • A corollary the paper does not pursue: the per-concept eigenvalue spectrum produced by the PCA is a quantitative measure of how much visual knowledge the model associates with a concept, so the same pipeline could rank concepts by richness or detect when finetuning or distillation erodes a model's knowledge.
  • Because the decomposition inherits the embedding's geometry, 'the model's visual capabilities' should be read as 'the capabilities visible through the chosen encoder'; the paper's own experiments switching between CLIP, DINO-v2, and FaceNet show that the discovered directions change with the encoder, so the method reveals the model through a lens rather than a single ground-truth structure.
  • The recipe — sample, embed, take principal components, attach low-rank adapters — is generic enough to carry over to other generative settings such as video or 3D models, where the same question of the main axes of variation for a prompt currently has no answer.
  • Since 10 PCA-based sliders match 64 attribute-specified ones on style coverage, the value of each additional slider depends on the embedding's geometry; an extension would use the PCA eigenvalue decay to choose the number of sliders per concept automatically instead of fixing a default.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes SliderSpace, a method that automatically decomposes the visual variation of a text-to-image diffusion model for a given prompt into a set of controllable directions. Given a prompt, the method generates a large set of samples from the model, encodes them with CLIP, and computes PCA components of the resulting embedding distribution. It then trains one low-rank adapter (LoRA) per principal component, with Eq. (5) maximizing the cosine similarity between the adapter's induced CLIP-embedding displacement and the corresponding principal axis. The adapters can be scaled and composed, and are evaluated on concept decomposition, artistic style exploration, and diversity enhancement for distilled models, with FID, DreamSim, CLIP score, and MTurk user studies. The paper also reports that directions transfer to related and out-of-domain concepts when trained with FaceNet embeddings.

Significance. If the central claims hold, SliderSpace is a practically useful tool: it provides unsupervised, composable control directions for diffusion models, works across U-Net and transformer architectures, and the paper ships code, data, and trained weights. The user studies and FID/DreamSim results provide credible aggregate evidence that random slider combinations increase diversity and stylistic coverage while preserving text alignment. The transferability experiment with FaceNet is a particularly interesting falsifiable result. However, the load-bearing claims that the discovered directions are individually 'human-understandable' and 'semantically orthogonal' are not quantitatively established. The method's decomposition is defined and evaluated in CLIP space, so the claim that it decomposes the model's visual knowledge is only as strong as the assumed alignment between CLIP geometry and human perception. The manuscript overstates the guarantee of semantic orthogonality given the displayed training objective, and the only per-slider interpretability evidence is generated by Claude-3.5 and is acknowledged by the authors to be imperfect.

major comments (4)
  1. [Section 4.2, Eq. (5)] The central 'semantic orthogonality' claim is not supported by the displayed objective. Eq. (5) minimizes 1 - cos(Δφ_i, v_i); maximizing cosine similarity to the target axis does not penalize large projections of Δφ_i onto v_j for j ≠ i. Therefore the statement in Section 4.2 that semantic orthogonality is 'guaranteed through the PCA-based initialization' is a non sequitur: the initial target axes are orthogonal, but nothing in the objective constrains the trained adapters' induced embedding displacements to remain within their assigned axes. Appendix E refers to a 'uniqueness criteria in Eqn 5' and a 'contrastive objective' that do not appear in the printed Eq. (5), suggesting either an omitted term or an unmet enforcement. Please report the post-training cross-cosine matrix cos(Δφ_i, v_j), or the pairwise matrix cos(Δφ_i, Δφ_j), for the discovered sliders. Without such a measurement, the claim of compositional control is not distinguished from mixed control.
  2. [Section 5.1 and Appendix D.3] The claim that individually discovered directions are 'human-understandable' is supported only by Claude-3.5-generated labels (Figs. 3 and E.14-E.19), and the paper itself states in Appendix D.3 that these captions are not always accurate. There is no quantitative test of per-slider semantic specificity, such as a forced-choice user evaluation in which participants identify the attribute changed by a single slider, or a retrieval/classification test of each slider's effect. The diversity metrics (DreamSim, FID) and user studies in Tables 1 and 2 measure aggregate diversity and usefulness of random slider combinations, not whether each slider has a distinct, interpretable identity. Please add such an evaluation or temper the 'human-understandable' claim accordingly.
  3. [Section 4.2, Eqs. (4)-(5)] The discovery pipeline is partly circular in a specific sense: the PCA axes are computed from CLIP embeddings of the model's samples, and each slider is then trained to reproduce exactly those axes in CLIP space. By construction the sliders will align with the principal modes of the CLIP embedding distribution; the substantive claim is that these modes correspond to the model's visual capabilities and to human perception. This is an empirical claim that should be tested, for example by comparing slider identity across encoders (CLIP vs. DINO-v2 vs. FaceNet), measuring whether a slider's effect transfers across encoders, and correlating slider labels with human judgments. As written, the conclusion that the directions 'decompose the visual structure of model's knowledge' is strictly a statement about CLIP-space geometry.
  4. [Tables 1-2 and Figure 4] The user-study win rates and the FID/DreamSim values are reported as point estimates without confidence intervals, number of participants, or significance tests. Since the user studies are the primary external validation of the diversity and usefulness claims, please provide per-task participant counts and 95% confidence intervals (or bootstrap intervals for FID). This is needed to assess whether the reported preferences and FID improvements are reliable.
minor comments (5)
  1. [Section 3.1] Eq. (2) uses \(\bar{\alpha}_t\) without defining it; please define \(\bar{\alpha}_t = \prod_{s=1}^t \alpha_s\) for clarity.
  2. [Section 5 and Appendix B.2] The main text reports discovery of 64 semantic directions, while Appendix B.2 states that the default configuration is 40 PCA directions with rank-one adapters; please reconcile these numbers and state the exact settings used for each experiment.
  3. [Figure 5] The caption lists panels as (a) Real Artists, (e) SliderSpace, (c) LLM Prompts, (b) Generic Prompts, (d) Concept Sliders, but the visual order of the figure should be reordered to match the caption labels.
  4. [Appendix E] The abbreviations in the Figure E.1 caption, such as 'w/o Contrast' and 'w/o CLIP', are not defined in the caption; please add a sentence explaining each ablation condition.
  5. [Abstract and Code Availability] The URL is given as sliderspace.baulab.info; please include the full https:// URL for consistency with the stated code and data release.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: SliderSpace's sliders are trained to realize PCA directions extracted from the model's own CLIP distribution, and the main claims are tested against external metrics.

full rationale

The derivation chain is self-contained rather than circular. Equation 4 computes PCA in CLIP space over final-image extrapolations from the target model, and Equation 5 trains each LoRA so that its CLIP-space displacement aligns with one principal component. This is a constructive fitting procedure: the PCA axes are summary statistics of the model's output distribution, and the trained adapters are new parameters whose value is assessed by external measures (DreamSim diversity, FID against artist-name distributions, user studies, and FaceNet transferability), none of which uses the fitted PCA axes as ground truth. The concerning sentence in Section 4.2, 'semantic orthogonality is guaranteed through the PCA-based initialization of directions as each principal components are mutually orthogonal,' is not established by the printed objective: Eq. 5 only maximizes cosine similarity to each v_i and contains no cross-term preventing a slider's displacement from projecting onto other PCA axes, and Appendix E refers to a 'uniqueness criteria' and 'contrastive objective' that do not appear in the printed Eq. 5. This is a missing proof or possibly an omitted term, and it is a correctness and reproducibility gap, but it is not an equation reducing to its own input. The self-citations, such as Concept Sliders [16] and StyleCLIP [33], supply the adapter and embedding-arithmetic machinery, but the PCA-based decomposition is specified in this paper and externally benchmarked, so the self-citations are not load-bearing. Overall, no step in the claimed derivation reduces by construction to the paper's inputs.

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

The method's central decomposition rests on CLIP as a semantic ground truth and on LoRA as a linear control mechanism; both are assumed without independent proof. Hyperparameters for direction count and rank are tuned by hand. No new physical entities are introduced.

free parameters (4)
  • number of PCA directions n = 40 (default for most experiments, up to 64)
    Chosen by hand; hyperparameter analysis shows diversity improves up to about 40 directions then plateaus; central decomposition depends on this cutoff.
  • LoRA rank = 1 (rank-one adapters)
    Chosen based on ablation showing rank-one best under fixed training budget; affects how faithfully each slider can reproduce its PCA component.
  • training sample count m = ~5000 for concept prompts, 8000 for COCO diversity
    Sampling budget for PCA; is manually set and affects stability of principal components.
  • sparse slider activation budget = 3 random sliders out of 32 or 64 during evaluation
    Evaluation protocol chooses 3 sliders randomly; diversity scores depend on this choice rather than a single controlled slider.
assumptions (5)
  • domain assumption PCA in CLIP embedding space identifies the principal modes of variation of the diffusion model's visual distribution for a prompt.
    Section 4.2 Eq. 4 uses phi(x0,t) from CLIP as the semantic space; if CLIP geometry does not match the model's visual variation, the discovered directions are not the model's principal visual directions.
  • domain assumption Adding a LoRA to cross-attention layers and maximizing cosine similarity in CLIP space induces a persistent, controllable change in generated images.
    Section 4.2 and Eq. 5 train adapters with this objective; the paper does not prove these weight updates are linear or stable across seeds beyond qualitative evidence.
  • domain assumption Final Image Extrapolation (Eq. 2) gives a clean-image estimate good enough for CLIP feature analysis at each timestep.
    Section 3.1 introduces Eq. 2; all CLIP features are computed on extrapolated latents/images, and errors in the estimate propagate into PCA.
  • domain assumption Semantic orthogonality of PCA components in CLIP space transfers to perceptual independence when combining sliders.
    Section 4.2 claims orthogonality guarantees each slider captures a distinct mode, but sliders are only orthogonal in CLIP space, not in image space or human perception.
  • standard math Standard PCA and cosine similarity are accepted math tools.
    Used in Eq. 4 and 5 without proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SliderSpace: Decomposing the Visual Capabilities of Diffusion Models." pith.science (2026). https://pith.science/paper/WRUHWO4A

@misc{pith2026250201639,
  author       = {Pith},
  title        = {Pith review of: SliderSpace: Decomposing the Visual Capabilities of Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WRUHWO4A}},
  note         = {Machine review of arXiv:2502.01639}
}
read the original abstract

We present SliderSpace, a framework for automatically decomposing the visual capabilities of diffusion models into controllable and human-understandable directions. Unlike existing control methods that require a user to specify attributes for each edit direction individually, SliderSpace discovers multiple interpretable and diverse directions simultaneously from a single text prompt. Each direction is trained as a low-rank adaptor, enabling compositional control and the discovery of surprising possibilities in the model's latent space. Through extensive experiments on state-of-the-art diffusion models, we demonstrate SliderSpace's effectiveness across three applications: concept decomposition, artistic style exploration, and diversity enhancement. Our quantitative evaluation shows that SliderSpace-discovered directions decompose the visual structure of model's knowledge effectively, offering insights into the latent capabilities encoded within diffusion models. User studies further validate that our method produces more diverse and useful variations compared to baselines. Our code, data and trained weights are available at https://sliderspace.baulab.info

Figures

Figures reproduced from arXiv: 2502.01639 by the authors.

Figure 1
Figure 1. Given a prompt, SliderSpace identifies the principal directions of the visual capabilites of a diffusion model by decomposing the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Given a prompt, SliderSpace first generates images and extracts the CLIP features. We then compute the spectral decompostion [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. SliderSpace decomposes the visual variation of diffusion model’s knowledge corresponding to a concept. These directions can [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: (left) SliderSpace generates diverse variations of a concept, as measured by DreamSim [14] distance across generated samples [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: SliderSpace demonstrates broader artistic style coverage, as evidenced by the lower FID scores compared to both supervised [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Comparison of artistic styles discovered in SDXL-DMD: (Left) Representative samples from artist-specific prompts manually [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: SliderSpace can also decompose general broad visual variation of diffusion model’s and can be used to overcome model collapse [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: SliderSpace directions for the “person” concept successfully generalize to related “police” and “athlete” concepts. They also [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Show Me Examples: Inferring Visual Concepts from Image Sets

    cs.CV 2026-07 unverdicted novelty 7.0 of 10

    Introduces VICIS task and training framework for inferring visual concepts from image sets, with experiments showing better accuracy, diversity, and generalization than standard VLMs on synthetic and ImageNet data.

  2. Kontinuous Kontext: Continuous Strength Control for Instruction-based Image Editing

    cs.CV 2025-10 conditional novelty 6.0 of 10

    Kontinuous Kontext adds continuous edit-strength control to instruction-based image editing by projecting a scalar strength and text embedding into the modulation space of a Flux Kontext diffusion editor.

Reference graph

Works this paper leans on

52 extracted references · 39 canonical work pages · cited by 2 Pith papers

  1. [1]

    Im- age2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF inter- national conference on computer vision, pages 4432– 4441, 2019

    Rameen Abdal, Yipeng Qin, and Peter Wonka. Im- age2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF inter- national conference on computer vision, pages 4432– 4441, 2019

  2. [2]

    Styleflow: Attribute-conditioned explo- ration of stylegan-generated images using conditional continuous normalizing flows

    Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Pe- ter Wonka. Styleflow: Attribute-conditioned explo- ration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1–21, 2021

  3. [3]

    Introducing claude 3.5 sonnet, 2024

    Anthropic. Introducing claude 3.5 sonnet, 2024

  4. [4]

    Announcing state-of-the-art flux.1 dev and schnell models, 2024

    BlackForestLabs. Announcing state-of-the-art flux.1 dev and schnell models, 2024

  5. [5]

    Ledits++: Limitless im- age editing using text-to-image models

    Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kerst- ing, and Apolin ´ario Passos. Ledits++: Limitless im- age editing using text-to-image models. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861–8870, 2024

  6. [6]

    Large scale gan training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv e-prints, pages arXiv–1809, 2018

  7. [7]

    Instructpix2pix: Learning to follow image edit- ing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image edit- ing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pages 18392–18402, 2023

  8. [8]

    Training-free regional prompt- ing for diffusion transformers

    Anthony Chen, Jianjin Xu, Wenzhao Zheng, Gaole Dai, Yida Wang, Renrui Zhang, Haofan Wang, and Shanghang Zhang. Training-free regional prompt- ing for diffusion transformers. arXiv preprint arXiv:2411.02395, 2024

Show all 52 references
  1. [9]

    Noiseclr: A con- trastive learning approach for unsupervised discovery of interpretable directions in diffusion models

    Yusuf Dalva and Pinar Yanardag. Noiseclr: A con- trastive learning approach for unsupervised discovery of interpretable directions in diffusion models. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24209–24218, 2024

  2. [10]

    Turboedit: Text-based image editing using few-step diffusion models, 2024

    Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024

  3. [11]

    Diffusion models beat gans on image synthesis

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

  4. [12]

    Interpreting the weight space of customized diffusion models

    Amil Dravid, Yossi Gandelsman, Kuan-Chieh Wang, Rameen Abdal, Gordon Wetzstein, Alexei A Efros, and Kfir Aberman. Interpreting the weight space of customized diffusion models. arXiv preprint arXiv:2406.09413, 2024

  5. [13]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Confer- ence on Machi...

  6. [14]

    Dreamsim: Learning new dimensions of hu- man visual similarity using synthetic data

    Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of hu- man visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024

  7. [15]

    Stylegan-nada: Clip-guided domain adaptation of im- age generators

    Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of im- age generators. ACM Transactions on Graphics (TOG), 41(4):1–13, 2022

  8. [16]

    Concept sliders: Lora adaptors for precise control in diffusion models

    Rohit Gandikota, Joanna Materzy ´nska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models. In European Conference on Computer Vision , pages 172–188. Springer, 2024

  9. [17]

    Generative adversar- ial nets

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

  10. [18]

    Towards a framework for human-ai interaction patterns in co-creative gan appli- cations

    Imke Grabe, Miguel Gonz’alez-Duque, Sebastian Risi, and Jichen Zhu. Towards a framework for human-ai interaction patterns in co-creative gan appli- cations. In Joint Proceedings of the ACM IUI Work- shops, 2022

  11. [19]

    Ganspace: Discovering inter- pretable gan controls

    Erik H ¨ark¨onen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering inter- pretable gan controls. Advances in neural information processing systems, 33:9841–9850, 2020

  12. [20]

    Prompt- to-prompt image editing with cross attention control

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

  13. [21]

    Style aligned image generation via 9 shared attention

    Amir Hertz, Andrey V oynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via 9 shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pages 4775–4785, 2024

  14. [22]

    Clipscore: A reference- free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ro- nan Le Bras, and Yejin Choi. Clipscore: A reference- free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021

  15. [23]

    On modeling human-computer co- creativity

    Oliver Hoffmann. On modeling human-computer co- creativity. In Knowledge, Information and Creativity Support Systems: Selected Papers from KICSS’2014- 9th International Conference, held in Limassol, Cyprus, on November 6-8, 2014 , pages 37–48. Springer, 2016

  16. [24]

    Lora: Low-rank adaptation of large language models

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

  17. [25]

    Image synthesis style studies, 2022

    Surea I, Proxima Centauri B, Erratica, and Stephen Young. Image synthesis style studies, 2022

  18. [26]

    Improving image generation with better captions

    et al James Betker. Improving image generation with better captions. OpenAI Reports, 2023

  19. [27]

    Scaling up gans for text-to-image synthesis

    Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124–10134, 2023

  20. [28]

    A style- based generator architecture for generative adversar- ial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style- based generator architecture for generative adversar- ial networks. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 4401–4410, 2019

  21. [29]

    Multi-concept cus- tomization of text-to-image diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept cus- tomization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931–1941, 2023

  22. [30]

    Photomaker: Cus- tomizing realistic human photos via stacked id embed- ding

    Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Cus- tomizing realistic human photos via stacked id embed- ding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024

  23. [31]

    Unsupervised compo- sitional concepts discovery with text-to-image genera- tive models

    Nan Liu, Yilun Du, Shuang Li, Joshua B Tenen- baum, and Antonio Torralba. Unsupervised compo- sitional concepts discovery with text-to-image genera- tive models. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision , pages 2085– 2095, 2023

  24. [32]

    Zero- shot image-to-image translation

    Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero- shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023

  25. [33]

    Styleclip: Text-driven manipulation of stylegan imagery

    Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2085–2094, 2021

  26. [34]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

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

  27. [35]

    Unsupervised representation learning with deep convolutional generative adversarial net- works

    Alec Radford. Unsupervised representation learning with deep convolutional generative adversarial net- works. arXiv preprint arXiv:1511.06434, 2015

  28. [36]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sas- try, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, ...

  29. [37]

    Stable diffusion 2.0 release, 2022

    Robin Rombach. Stable diffusion 2.0 release, 2022

  30. [38]

    Stable diffusion v1-4 model card, 2022

    Robin Rombach and Patrick Esser. Stable diffusion v1-4 model card, 2022

  31. [39]

    Stable diffusion v2 model card, 2022

    Robin Rombach and Patrick Esser. Stable diffusion v2 model card, 2022

  32. [40]

    High- resolution image synthesis with latent diffusion mod- els

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ˜A¶rn Ommer. High- resolution image synthesis with latent diffusion mod- els. In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), 2022

  33. [41]

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

  34. [42]

    Cloneofsimo/lora: Using low-rank adapta- tion to quickly fine-tune diffusion models.s

    Simo Ryu. Cloneofsimo/lora: Using low-rank adapta- tion to quickly fine-tune diffusion models.s. GitHub, 2023

  35. [43]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87–

  36. [44]

    Facenet: A unified embedding for face recog- nition and clustering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recog- nition and clustering. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 815–823, 2015

  37. [45]

    Interfacegan: Interpreting the disentangled face representation learned by gans

    Yujun Shen, Ceyuan Yang, Xiaoou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence , 44(4): 2004–2018, 2020. 10

  38. [46]

    Instantbooth: Personalized text-to-image generation without test-time finetuning

    Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 8543–8552, 2024

  39. [47]

    Stylespace analysis: Disentangled controls for style- gan image generation

    Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for style- gan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 12863–12872, 2021

  40. [48]

    Turboedit: Instant text-based image editing

    Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing. ECCV, 2024

  41. [49]

    Inversion-free image editing with natu- ral language

    Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natu- ral language. In Conference on Computer Vision and Pattern Recognition 2024, 2024

  42. [50]

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023

  43. [51]

    Improved distribution matching distil- lation for fast image synthesis

    Tianwei Yin, Micha ¨el Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distil- lation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024

  44. [52]

    person” show higher varia- tion in CLIP space compared to rarer concepts like “waterfalls

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision , pages 3836– 3847, 2023. 11 SliderSpace: Decomposing the Visual Capabilities of Diffusion Mo...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.