Pith. sign in

REVIEW 3 major objections 6 minor 48 references

This paper claims that artists can learn to predict a text-to-image diffusion model's visual behavior by bending its internal layers, turning a black box into a creative material.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 04:44 UTC pith:AFIY3IO6

load-bearing objection A genuinely useful toolkit and an honest preliminary mapping, but the abstract's 'consistent families' claim outruns the statistics; worth refereeing with revisions. the 3 major comments →

arxiv 2607.22428 v1 pith:AFIY3IO6 submitted 2026-07-24 cs.HC cs.AIcs.LGcs.MM

Unboxing Diffusion Models for the Arts: Interactive Model Bending and Practice-Based Explainability

Ahmed M. Abuzuraiq , Philippe Pasquier This is my paper
classification cs.HC cs.AIcs.LGcs.MM MSC 68T0768T45
keywords model bendingdiffusion modelsStable Diffusion 1.5explainable AIcreative practicelatent diffusioninteractive interfacespractice-based XAI
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that explainability for artists is not about post-hoc explanations but about hands-on intervention: making a large generative model's internal structure visible and manipulable. To support this, the authors build a model-bending toolkit inside ComfyUI and an interactive web interface for selecting UNet layers and applying operations like scaling, ablation, rotation, and noise. Through systematic experiments on Stable Diffusion 1.5, they show that bending a given layer produces relatively consistent visual effects across different seeds and prompts, with layer type and magnitude mattering more than spatial location. If true, artists can develop practical heuristics for what to bend, when, and by how much, gaining craft-like familiarity with a model without needing mechanistic understanding.

Core claim

The central discovery is that intervening on specific components of the SD1.5 UNet produces consistent families of visual effects, enabling practice-based explainability. Quantitatively, the input region of the UNet has a stronger impact on latent divergence than the output region (mean cosine distance 0.1776 vs 0.1088), Residual Blocks are more impactful than Cross-Attention layers (0.1333 vs 0.1070), and normalization layers are as impactful as convolutional layers. Layer type exerts a stronger influence than spatial location, contrasting with StyleGAN findings, and bending during early denoising steps predominantly shapes global structure while later steps yield fine-grained changes. Thes

What carries the argument

The key mechanism is the bending operator: a training-free intervention that hooks into a specified layer's output during inference and transforms the tensor using operations like scalar multiplication, ablation, rotation, or noise. Layer paths are specified via the UNet's hierarchical structure (e.g., 'diffusion model.middle block.0.in layers'), and the system copies the model definition so bends affect only downstream parts of the workflow. The interactive interface adds an SVG layer explorer and sliders, letting artists select and bend layers without scripting.

Load-bearing premise

The hand-picked layers in the study are representative of the UNet regions and module types they illustrate, so the reported hierarchy reflects architectural properties rather than a biased selection of layers.

What would settle it

Run a systematic ablation sweep across all UNet layers (not a hand-picked subset) with multiple seeds and a fixed prompt, then group by layer type and location. If the variance in latent distance within a layer type exceeds the variance between types, or if input-block and output-block means overlap substantially, the claimed hierarchy collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Artists can build model-specific cheat sheets mapping layers, magnitudes, and timestep ranges to visual outcomes, making bending reproducible and shareable.
  • Bending effects that are consistent across seeds and prompts allow artists to achieve intended stylistic shifts without prompting tricks.
  • The finding that early timesteps dominate global structure lets artists control coarse vs fine changes by limiting the denoising step range.
  • The tools lower the barrier for non-technical artists to engage with model internals, supporting AI literacy and critical engagement.
  • The approach can extend beyond the UNet to other pipeline components, such as VAE, CLIP embeddings, and LoRA matrices, broadening creative control.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The layer-type-over-location hierarchy may reflect a general property of iterative denoising, suggesting similar patterns could emerge in newer diffusion transformers, though this remains untested.
  • A systematic sweep across all layers of each type, rather than hand-picked examples, could confirm whether the quantitative hierarchy holds or is an artifact of selection.
  • The consistency of bending effects raises the possibility of using bending as a probe for model robustness or safety guardrails, as the ethics statement itself hints.
  • Perceptual metrics like LPIPS or DINOv2, which the paper suggests but does not run, might reveal whether latent cosine distance aligns with human-perceived visual impact.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. This paper argues that large text-to-image diffusion models can be treated as creative materials when their internal structure is exposed and manipulable, and that practice-based explainability through 'model bending' can serve artists. The authors contribute a ComfyUI plugin and an interactive web interface for inspecting Stable Diffusion 1.5's UNet and applying training-free interventions at selected layers and timesteps. They report a systematic study that enumerates bending locations, seeds, prompts, and timestep ranges, using qualitative image grids and latent-space cosine distance from the unbent baseline. The paper's central claim is that bending specific components yields relatively consistent families of visual effects, enabling artists to develop layer-level heuristics for what to bend, when, and by how much.

Significance. If the central claim holds, the paper offers a concrete path toward 'doing-based' XAI for the arts and a practical toolkit for artists. The system contribution is genuine: a reproducible, open-source implementation integrated into a widely used workflow, with interactive layer selection and caching of experimental results. The study design is transparent, and the authors' limitation section candidly acknowledges the lack of perceptual and human evaluation. The main weakness is that the consistency claim—the core of the paper—is not supported by the quantitative analysis as currently presented. Because that claim is testable with existing data, the manuscript is promising but needs additional evidence.

major comments (3)
  1. [Section 4.4 / Abstract] The paper's central claim—that bending specific components produces 'relatively consistent families of visual effects'—is not directly tested. The only quantitative evidence in Section 4.3 is cosine distance from the unbent latent, which measures the magnitude of the intervention, not whether the same layer produces a consistent visual character across seeds/prompts or whether different layers produce distinguishable effects. The qualitative grids (Figs. 8–17) are suggestive but are not a substitute. I ask the authors to add a pairwise similarity analysis: at matched bending magnitude, compare same-layer outputs across seeds/prompts against different-layer outputs, using a perceptual or latent metric (e.g., LPIPS, cosine distance), with a permutation or bootstrap test. This would substantiate or refine the hierarchy claimed in Section 4.4.
  2. [Section 4.3] The text states 'significant disparities in impact' and 'a clear hierarchy where Input>Output and ResBlocks>Cross-Attention,' but no inferential statistics are reported. The descriptive means and standard deviations overlap widely (e.g., LayerNorm 0.1688±0.1562 vs Conv2d 0.1685±0.1658; ResBlocks 0.1333±0.1541 vs Cross-Attention 0.1070±0.1110). Please provide confidence intervals, effect sizes, and tests (permutation or bootstrap) for the stated differences, or revise the language to 'observed' rather than 'significant.'
  3. [Section 4.2.2] The cross-seed and cross-prompt experiments use 'select layers, hand-picked to span input, middle and output blocks' (five layers). No criterion for representativeness is given, and the quantitative hierarchy in Section 4.3 is computed over these and other selected layers. This selection could bias the conclusions about layer type and location. Please justify the selection or sample layers systematically; alternatively, explicitly restrict the generalization claims to the selected layers.
minor comments (6)
  1. [Page 9] Typo: 'the scripts used to generate the results in the in Section 4' should read 'in Section 4.'
  2. [Section 4.2.3] The opening sentence says 'Using the same prompt' but the experiment varies prompts; this should say 'Using the same seed.'
  3. [Section 3.3] The phrase 'We address goals' is missing 'these'; consider reformulating.
  4. [Section 4.2.2/Figures 8–12] Layer path labels are inconsistent (e.g., 'time embed.2' vs 'time embed.2' in text; 'input blocks.1.0 .in layers.0' with spaces). Please standardize the notation.
  5. [Section 2.2] T5 is described as a contrastive text encoder; T5 is not trained contrastively. Please clarify or correct.
  6. [Figure 1] Color coding ('gold' vs 'purple') may be difficult for color-blind readers; add labels or patterns.

Circularity Check

0 steps flagged

No circularity: the 'relatively consistent families' claim is an empirical summary of generated outputs, not a consequence of fitted inputs or self-cited assumptions.

full rationale

The paper's central claim is that bending specific components of SD1.5 produces relatively consistent visual families. This is an empirical summary of generated grids (Figs. 6–17) and a post-hoc cosine-distance metric, not a result derived from equations that already contain the conclusion. No parameters are fitted and then re-predicted; the bending operators, layer paths, and baseline latents are all inputs, and the distance-to-baseline is an independent measurement. The quantitative analysis only summarizes distances by architectural grouping, so the hierarchy (Input>Output, ResBlocks>Cross-Attention) is descriptive rather than circular. The self-citations ([9], [13], [14]) support background framing and prior tools, but are not used as a uniqueness theorem or as the source of the empirical hierarchy. The paper itself restricts its scope: 'our intent is not to establish findings that generalize across different UNet architectures' (Sec. 5.1) and 'observations reported here should not be assumed to generalize' (Sec. 6). The main concerns—hand-picked layers and the absence of a statistical test of cross-seed consistency—are validity/selection issues, not circularity. Score 1 reflects only the presence of minor, non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The paper makes no mathematical derivation and fits no parameters; its claims rest on methodological assumptions about the metric, the representativeness of hand-picked layers, and the reliability of visual inspection.

axioms (3)
  • domain assumption Cosine distance between final latent representations is a meaningful proxy for the visual impact of a bending intervention.
    Section 4.1 defines the quantitative metric; Section 6 acknowledges it 'does not fully capture perceptual, aesthetic, or semantic aspects'.
  • ad hoc to paper The hand-picked layers in the cross-seed and cross-prompt experiments are representative of the UNet regions and module types they are chosen to span.
    Section 4.2.2 describes layers as 'hand-picked to span input, middle and output blocks', but gives no representativeness justification; the Input>Output hierarchy (Section 4.3) could depend on this selection.
  • domain assumption Qualitative visual comparison of image grids reliably detects consistent families of visual effects.
    Section 4.4 bases the 'relatively consistent visual impacts' claim on the authors' inspection of Figures 6-17 without inter-rater reliability or perceptual metrics.

pith-pipeline@v1.3.0-alltime-deepseek · 14749 in / 13138 out tokens · 132624 ms · 2026-08-01T04:44:50.276904+00:00 · methodology

0 comments
read the original abstract

Explainable AI (XAI) in creative practice can be less about technocentric explanation and more about enabling artists to inspect modify and debug models as part of making Yet largescale texttoimage diffusion systems are typically presented as opaque endtoend tools limiting this kind of material engagement We argue that even large models can function as creative materials when their internal structure is made visible and manipulable To support this we propose a handson approach to explainability centred on experimentation and intervention We instantiate this approach with a model bending and an interactive (inspection) interface integrated into ComfyUIs nodebased workflow including interactive layer selection and intervention controls Through qualitative and quantitative analysis of bending interventions in Stable Diffusion 15 we show how manipulating specific components of a diffusion pipeline produces relatively consistent families of visual effects allowing artists to build practical layerlevel intuition about how different parts of the model shape generated images

Figures

Figures reproduced from arXiv: 2607.22428 by Ahmed M. Abuzuraiq, Philippe Pasquier.

Figure 1
Figure 1. Figure 1: Basic workflow in ComfyUI (gold). The UNet bending nodes are shown in [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: JSON description of bending operations. 3.4 Interactive Bending Considering the complexity and scale of modern diffusion models, as well as the wide range of bending parameters, we found it important to support users in interactively identifying the parts they wish bend. We describe these supports next. 3.4.1 Model Inspection For more granular control, including the ability to select specific blocks or lay… view at source ↗
Figure 3
Figure 3. Figure 3: With the model inspector node, artists can interactively pick layers from the [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Interactive Interface for Bending Diffusion Models. The interface shows the [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: In each UNet block, users can drill down into one of the containers (higher [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: One prompt; rows: bending layer path, columns: module [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: One prompt; rows: bending layer path, columns: module [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Bending time embed.2; Same prompt, different seeds [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Bending input blocks.1.0.out layers.2; Same prompt, different seeds. Multiply by / Seed 42 0 123 456 789 786 0.0 0.5 1.0 1.5 2.0 [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Bending input blocks.4.0.skip connection; Same prompt, different seeds [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Bending middle block.1.transformer blocks.0.attn2.to out; Different seeds. Multiply by / Seed 42 0 123 456 789 786 0.0 0.5 1.0 1.5 2.0 [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Bending output blocks.3.1.norm; Same prompt, different seeds [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Results of bending the layer time embed.2; Different prompts, same seed [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Results of bending the layerinput blocks.1.0.out layers.2; Different prompts, same seed. Multiply Prompt 1 Prompt 2 Prompt 3 Prompt 4 Prompt 5 Prompt 6 0.0 0.5 default 1.5 2.0 [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Results of bending the layer input blocks.4.0.skip connection; Different prompts, same seed [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Bending middle block.1.transformer blocks.0.attn2.to out.0; Different prompts, same seed. Multiply Prompt 1 Prompt 2 Prompt 3 Prompt 4 Prompt 5 Prompt 6 0.0 0.5 default 1.5 2.0 [PITH_FULL_IMAGE:figures/full_fig_p020_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Bending the layer output blocks.3.1.norm; Different prompts, same seed [PITH_FULL_IMAGE:figures/full_fig_p020_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Varying the denoising time steps (out of 20) during which bending is applied. [PITH_FULL_IMAGE:figures/full_fig_p021_18.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    XAI—Explainable artificial intelligence

    David Gunning et al. “XAI—Explainable artificial intelligence”. In:Science robotics4.37 (2019)

    2019

  2. [2]

    Exploring XAI for the Arts: Explaining Latent Space in Generative Music

    Nick Bryan-Kinns et al. “Exploring XAI for the Arts: Explaining Latent Space in Generative Music”. In:1st Workshop on eXplainable AI approaches for debugging and diagnosis. 2023

    2023

  3. [3]

    Explainability Paths for Sustained Artistic Practice with AI

    Austin Tecks, Thomas Peschlow, and Gabriel Vigliensoni. “Explainability Paths for Sustained Artistic Practice with AI”. In:In Proceedings of Ex- plainable AI for the Arts Workshop 2024 (XAIxArts 2024). arXiv, July 21, 2024.doi:10.48550/arXiv.2407.15216. arXiv:2407.15216 [cs].url: http://arxiv.org/abs/2407.15216(visited on 04/08/2025). 28 Ahmed M. Abuzurai...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2407.15216 2024

  4. [4]

    XAIxArts Manifesto: Explainable AI for the Arts

    Nick Bryan-Kinns et al. “XAIxArts Manifesto: Explainable AI for the Arts”. In:Extended Abstracts of the CHI Conference on Human Factors in Com- puting Systems. CHI. Yokohama, Japan, 2025.doi:10 . 1145 / 3706599 . 3716227. arXiv:2502.21220 [cs].url:http://arxiv.org/abs/2502. 21220(visited on 04/13/2025)

    Pith/arXiv arXiv 2025

  5. [5]

    A Small-Data Mindset for Generative AI Creative Work

    Gabriel Vigliensoni, Phoenix Perry, and Rebecca Fiebrink. “A Small-Data Mindset for Generative AI Creative Work”. In:In Generative AI in HCI Workshop, CHI ’22. New York, NY, USA, 2022, p. 5

    2022

  6. [6]

    ”Explorers of Unknown Planets

    Baptiste Caramiaux and Sarah Fdili Alaoui. “”Explorers of Unknown Planets” Practices and Politics of Artificial Intelligence in Visual Arts”. In:Proceedings of the ACM on Human-Computer Interaction6.CSCW2 (2022), pp. 1–24

    2022

  7. [7]

    Using Generative AI as an Artistic Material: A Hacker’s Guide

    Terence Broad. “Using Generative AI as an Artistic Material: A Hacker’s Guide”. In:In Proceedings of Explainable AI for the Arts Workshop 2024. 2024

    2024

  8. [8]

    How Artists Use AI as a Responsive Material for Art Creation

    Sonja Rozental, Michel Van Dartel, and Alwin De Rooij. “How Artists Use AI as a Responsive Material for Art Creation”. In: ISEA 2025: 30th International Symposium on Electronic/Emerging Art. Korea, 2025.doi:10.31234/osf. io/gjdnw.url:https://osf.io/gjdnw(visited on 06/10/2025)

    doi:10.31234/osf 2025

  9. [9]

    Seizing the Means of Produc- tion: Exploring the Landscape of Crafting, Adapting and Navigating Gener- ative AI Models

    Ahmed M. Abuzuraiq and Philippe Pasquier. “Seizing the Means of Produc- tion: Exploring the Landscape of Crafting, Adapting and Navigating Gener- ative AI Models”. In:the 3rd Generative AI and HCI Workshop. CHI ’24. Hawaii, USA, 2024

    2024

  10. [10]

    Network Bend- ing: Expressive Manipulation of Deep Generative Models

    Terence Broad, Frederic Fol Leymarie, and Mick Grierson. “Network Bend- ing: Expressive Manipulation of Deep Generative Models”. In:Artificial In- telligence in Music, Sound, Art and Design: 10th International Conference, EvoMUSART 2021, Held as Part of EvoStar 2021, Virtual Event, April 7–9, 2021, Proceedings 10. Springer, 2021

    2021

  11. [11]

    com / comfyanonymous / ComfyUI

    comfyanonymous.https : / / github . com / comfyanonymous / ComfyUI

  12. [12]

    John Wiley & Sons, 2005

    Reed Ghazala.Circuit-Bending: Build Your Own Alien Instruments. John Wiley & Sons, 2005

    2005

  13. [13]

    Autolume-Live: Turning GANs into a Live VJing Tool

    Jonas Kraasch and Philippe Pasquier. “Autolume-Live: Turning GANs into a Live VJing Tool”. In:Proceedings of the 10th Conference on Computation, Communication, Aesthetics & X. Coimbra, Portugal, 2022, pp. 152–169

    2022

  14. [14]

    https : / / www

    Metacreation Lab.Autolume: A Neural-network Based Visual Synthesizer. https : / / www . metacreation . net / autolume. 2024. (Visited on 02/13/2024)

    2024

  15. [15]

    Controlling the Image Generation Process with Parametric Acti- vation Functions

    Ilia Pavlov. “Controlling the Image Generation Process with Parametric Acti- vation Functions”. In:International Conference on Computational Creativity ICCC25. Brazil, 2025

    2025

  16. [16]

    Network Bend- ing of Diffusion Models for Audio-Visual Generation

    Luke Dzwonczyk, Carmine Emanuele Cella, and David Ban. “Network Bend- ing of Diffusion Models for Audio-Visual Generation”. In: 27th International Conference on Digital Audio Effects (DAFx24). arXiv, 2024-06-28.doi: 10.48550/arXiv.2406.19589. arXiv:2406.19589 [cs].url:http: //arxiv.org/abs/2406.19589(visited on 02/20/2025). Unboxing Diffusion Models for t...

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.19589 2024

  17. [17]

    2025.url: https://garincurtis.com/projects/beyond-prompts

    Garin Curtis.Beyond Prompts: Developing An Expressive Tool for Real-Time Manipulation of Diffusion Models through Active Divergence. 2025.url: https://garincurtis.com/projects/beyond-prompts

    2025

  18. [18]

    Patch Explorer: Interpreting Diffusion Models through Interaction

    Imke Grabe et al. “Patch Explorer: Interpreting Diffusion Models through Interaction”. In:Mechanistic Interpretability for Vision at CVPR 2025 (Non- proceedings Track). Mar. 31, 2025.url:https : / / openreview . net / forum?id=0n9wqVyHas(visited on 07/02/2025)

    2025

  19. [19]

    Mingi Kwon, Jaeseok Jeong, and Youngjung Uh.Diffusion Models Already Have a Semantic Latent Space. Mar. 29, 2023.doi:10.48550/arXiv.2210. 10960. arXiv:2210.10960 [cs].url:http://arxiv.org/abs/2210. 10960(visited on 05/07/2025). Pre-published

    Pith/arXiv arXiv doi:10.48550/arxiv.2210.10960 2023

  20. [20]

    Unstable Audio: Code Bending Text-to-Music Generation

    Nick Collins. “Unstable Audio: Code Bending Text-to-Music Generation”. In:Proceedings of the AES International Conference on Machine Learn- ing and Artificial Intelligence for Audio. 2025.url:https : / / durham - repository.worktribe.com/output/4280622(visited on 11/20/2025)

    arXiv 2025

  21. [21]

    Studio Report: Sound Synthe- sis with DDSP and Network Bending Techniques

    Matthew Yee-King and Louis McCallum. “Studio Report: Sound Synthe- sis with DDSP and Network Bending Techniques”. In:Proceedings of 2nd Conference on AI Music Creativity. AIMC. 2021

    2021

  22. [22]

    Jaden Fiotto-Kaufman et al.NNsight and NDIF: Democratizing Access to Open-Weight Foundation Model Internals. Apr. 1, 2025.doi:10 . 48550 / arXiv.2407.14561. arXiv:2407.14561 [cs].url:http://arxiv.org/ abs/2407.14561(visited on 02/12/2026). Pre-published

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.14561 2025

  23. [23]

    Pyvene: A Library for Understanding and Improving PyTorch Models via Interventions

    Zhengxuan Wu et al. “Pyvene: A Library for Understanding and Improving PyTorch Models via Interventions”. In:Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguis- tics: Human Language Technologies (Volume 3: System Demonstrations). Ed. by Kai-Wei Chang, Annie Lee, and Nazneen Rajani. Mexico City, Me...

    doi:10.18653/v1/2024.naacl- 2024

  24. [24]

    2022.url:https://github.com/davidbau/baukit

    David Bau.Baukit. 2022.url:https://github.com/davidbau/baukit

    2022

  25. [25]

    Hacking Generative Models with Differentiable Network Bending

    Giacomo Aldegheri et al. “Hacking Generative Models with Differentiable Network Bending”. 2023. arXiv:2310 . 04816.url:https : / / github . com/GAldegheri/gan-bending

    2023

  26. [26]

    Autolume: A GAN-based No-Coding Small Data and Model Crafting Visual Synthesizer for Artistic Creation

    Arshia Sobhan et al. “Autolume: A GAN-based No-Coding Small Data and Model Crafting Visual Synthesizer for Artistic Creation”. In:Proceedings of the Conference on Animation and Interactive Art. Expanded ’25. New York, NY, USA: Association for Computing Machinery, Nov. 5, 2025, pp. 165– 172.isbn: 979-8-4007-1532-7.doi:10 . 1145 / 3749893 . 3749968.url: htt...

    2025

  27. [27]

    Hidden Layer Interaction: A Technique to Explore the Material of Generative AI

    Imke Grabe and Tom Jenkins. “Hidden Layer Interaction: A Technique to Explore the Material of Generative AI”. In:Proceedings of the 2025 ACM Designing Interactive Systems Conference. DIS ’25. New York, NY, USA: Association for Computing Machinery, July 4, 2025, pp. 1913–1927.isbn: 30 Ahmed M. Abuzuraiq and Philippe Pasquier 979-8-4007-1485-6.doi:10.1145/3...

    arXiv 2025

  28. [28]

    High-Resolution Image Synthesis With Latent Dif- fusion Models

    Robin Rombach et al. “High-Resolution Image Synthesis With Latent Dif- fusion Models”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022, pp. 10684–10695.url:https : / / openaccess.thecvf.com/content/CVPR2022/html/Rombach_High- Resolution_Image_ Synthesis_With_Latent_Diffusion _Models_ CVPR_2022_paper(visited on 0...

    2022

  29. [29]

    Automatic1111 WebUI.https://github.com/AUTOMATIC1111/stable- diffusion-webui. 2024. (Visited on 02/13/2024)

    2024

  30. [30]

    Streamdiffusion: A pipeline-level solution for real-time interactive generation

    Akio Kodaira et al. “Streamdiffusion: A pipeline-level solution for real-time interactive generation”. In:Proceedings of the IEEE/CVF International Con- ference on Computer Vision. 2025, pp. 12371–12380

    2025

  31. [31]

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric

    Richard Zhang et al. “The Unreasonable Effectiveness of Deep Features as a Perceptual Metric”. In:CVPR. 2018

    2018

  32. [32]

    Dinov2: Learning robust visual features without super- vision

    Maxime Oquab et al. “Dinov2: Learning robust visual features without super- vision”. In:arXiv preprint arXiv:2304.07193(2023)

    Pith/arXiv arXiv 2023

  33. [33]

    Learning transferable visual models from natural lan- guage supervision

    Alec Radford et al. “Learning transferable visual models from natural lan- guage supervision”. In:International conference on machine learning. PmLR. 2021, pp. 8748–8763

    2021

  34. [34]

    Visual Parameter Space Exploration for AI Art Design

    Julian Hummel. “Visual Parameter Space Exploration for AI Art Design”. B.S. thesis. 2023

    2023

  35. [35]

    Visual Parameter Space Analysis: A Conceptual Framework

    Michael Sedlmair et al. “Visual Parameter Space Analysis: A Conceptual Framework”. In:IEEE Transactions on Visualization and Computer Graphics 20.12 (Dec. 2014), pp. 2161–2170.issn: 1077-2626.doi:10.1109/TVCG. 2014.2346321

    arXiv 2014

  36. [36]

    Fashioning Creative Expertise with Generative AI: Graphical Interfaces for Design Space Exploration Better Support Ideation than Text Prompts

    Richard Lee Davis et al. “Fashioning Creative Expertise with Generative AI: Graphical Interfaces for Design Space Exploration Better Support Ideation than Text Prompts”. In:Proceedings of the 2024 CHI Conference on Hu- man Factors in Computing Systems. Chi ’24. New York, NY, USA: As- sociation for Computing Machinery, 2024.isbn: 979-8-4007-0330-0.doi: 10 ...

    arXiv 2024

  37. [37]

    Generative Rotoscoping: A First-Person Autobiographical Ex- ploration on Generative Video-to-Video Practices

    David Ledo. “Generative Rotoscoping: A First-Person Autobiographical Ex- ploration on Generative Video-to-Video Practices”. In:Proceedings of the 2025 Conference on Creativity and Cognition. C&C ’25. New York, NY, USA: Association for Computing Machinery, June 22, 2025, pp. 931– 948.isbn: 979-8-4007-1289-0.doi:10 . 1145 / 3698061 . 3726926.url: https ...

    2025

  38. [40]

    Lora: Low-rank Adaptation of Large Language Models

    Edward J Hu et al. “Lora: Low-rank Adaptation of Large Language Models”

  39. [41]

    StabilityAI.Stable Diffusion.https://stability.ai/stable- image

  40. [42]

    (Visited on 08/30/2024)

    2024

  41. [43]

    On the commoditization of artificial intelligence

    Abdullah A Abonamah, Muhammad Usman Tariq, and Samar Shilbayeh. “On the commoditization of artificial intelligence”. In:Frontiers in psychology12 (2021)

    2021

  42. [44]

    A Taxonomy of the Biases of the Images created by Generative Artificial Intelligence

    Adriana Fern ´andez de Caleya V ´azquez and Eduardo C Garrido-Merch ´an. “A Taxonomy of the Biases of the Images created by Generative Artificial Intelligence”. In:Current Trends in Business Management 3(2024)

    2024

  43. [45]

    The Impact of Generative AI on Artists

    Reishiro Kawakami and Sukrit Venkatagiri. “The Impact of Generative AI on Artists”. In:Proceedings of the 16th Conference on Creativity & Cognition. 2024, pp. 79–82

    2024

  44. [46]

    Towards Directive Explanations: Crafting Explainable AI Systems for Actionable Human-AI Interactions

    Aditya Bhattacharya. “Towards Directive Explanations: Crafting Explainable AI Systems for Actionable Human-AI Interactions”. In:Extended Abstracts of the CHI Conference on Human Factors in Computing Systems. 2024, pp. 1–6

    2024

  45. [47]

    Active Divergence with Generative Deep Learning–A Survey and Taxonomy

    Terence Broad et al. “Active Divergence with Generative Deep Learning–A Survey and Taxonomy”. In:Proceedings of the 12th International Conference on Computational Creativity (ICCC ’21). 2021.isbn: 978-989-54-1603-5. doi:arXivpreprintarXiv:2107.05599

    Pith/arXiv arXiv 2021

  46. [48]

    AI in the public eye: Investigating public AI literacy through AI art

    Drew Hemment et al. “AI in the public eye: Investigating public AI literacy through AI art”. In:Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency. 2023, pp. 931–942

    2023

  47. [49]

    Long Ouyang et al.Training Language Models to Follow Instructions with Human Feedback. Mar. 4, 2022.doi:10.48550/arXiv.2203.02155. arXiv: 2203.02155 [cs].url:http://arxiv.org/abs/2203.02155(visited on 11/28/2025). Pre-published

    Pith/arXiv arXiv doi:10.48550/arxiv.2203.02155 2022

  48. [2024]

    (Visited on 02/13/2024)

    2024