Kunster: Artist Attribution and AI

• Kunster is a multifaceted term that denotes both artist attribution in computational recognition tasks and the active role of artists in scientific collaborations.
• It underpins a multi-task learning framework that integrates artist, material, and period attributes, thereby boosting recognition accuracy while reducing training redundancy.
• Kunster also inspires artist-centred AI systems and mobile neural style transfer applications, emphasizing unexpectedness, diversification, and user accessibility.

“Kunster” is used in contemporary research literature in several distinct but related senses. It denotes the artist as a target of computational recognition in the OmniArt framework, where artist attribution is learned jointly with artwork type, material, and creation period (Strezoski et al., 2017). It denotes the artist as an embedded collaborator in scientific settings, as in the Padova eXtreme19 project and the South Pole collaboration around IceCube (Prandini et al., 2021, Fortescue et al., 2019). It also appears as the proper name of “Kunster -- AR Art Video Maker,” a mobile system for real-time neural style transfer (Dudzik et al., 2020). A later conceptual synthesis, “Toward an Artist-Centred AI,” shifts attention from model-centric evaluation to the values of unexpectedness, diversification, affordance, and evolvability (Krekovic et al., 2024).

1. Semantic range and usage

In the cited literature, “Kunster” is not restricted to a single technical object. In OmniArt, it appears in the explicit formulation of a recognition task: the framework was designed to recognize artists (“Kunster”) while simultaneously predicting an artwork’s type, material and creation period. In the South Pole collaboration, the term is used directly for the human artistic agent: “the artist (“Kunster,” in our case Prof. Donald Fortescue).” In the Padova case, “This Kunster (‘artist’ in German) collaboration” names a collaboration between astrophysicists and high-school art students. In the mobile style-transfer paper, “Kunster” is the name of an application available in the Apple’s App Store (Strezoski et al., 2017, Fortescue et al., 2019, Prandini et al., 2021, Dudzik et al., 2020).

These usages suggest a heterogeneous vocabulary in which the same term can refer to at least three layers of practice: artist attribution, artist participation, and artist-facing software. That heterogeneity is significant because it prevents reduction of the topic to a single question such as authorship detection or image generation alone. Within the cited works, “Kunster” is instead attached to problems of representation, collaboration, interface design, and deployment.

2. Artist attribution as a multi-task learning problem

In OmniArt, artist recognition is formulated as one component of a multi-task architecture built around a single convolutional backbone whose parameters remain frozen; in the reported experiments this backbone is the 50-layer Residual Network pre-trained on ImageNet. The base feature extractor outputs a 2,048-dimensional feature vector for each image. Above it sits a shared fully connected “representation” layer, for which 6,144 hidden units with ReLU activation were found to offer the best trade-off, followed by separate lightweight heads for artist attribution, creation period, material, and type (Strezoski et al., 2017).

The task heads are heterogeneous. Artist attribution uses a softmax layer over $N_{\text{artist}}$ classes, with categorical cross-entropy. Creation period uses a single linear unit trained with mean absolute error. Material and type are multi-label heads with sigmoid activation and binary cross-entropy. These are combined through

$L_{\text{total}} = \sum_{t=1}^{T} w_t s_t L_t.$

For period regression, $s_{\text{period}}$ was set to $1/10$, or scale $= 0.1$, to prevent the typically larger regression loss from overwhelming classification gradients. The $w_t$ weights were tuned on the validation split to reflect relative importance and imbalance.

The dataset composition is correspondingly large and structured. The OmniArt Challenge contains 432,217 high-quality photographic reproductions with 21,364 distinct artists, 837 artwork types, 6,385 materials, and 2,389 period labels. The train/val/test split is 70%/20%/10%, done per artist so that every artist appears in each split in roughly the same proportion. Images were resized to $224 \times 224$, and only random horizontal flips were applied. The authors report that subtracting per-channel means or applying ZCA whitening degraded fine texture and brush-stroke detail important for artist attribution.

Training freezes all ResNet-50 convolutional layers and updates only the shared representation layer and the four task heads, keeping the number of trainable parameters under three million. The batch size is 32; optimization uses Adam with an initial learning rate of $10^{-4}$, decayed by $0.5$ every five epochs if validation loss stalled. The balanced loss was arranged so that artist attribution contributed about 40%, period regression 20%, type 20%, and material 20% of the gradient magnitude after applying $s_t$.

Quantitatively, the multi-task model outperforms single-task CNNs on four increasingly fine-grained artist splits. On the largest split, artists with $L_{\text{total}} = \sum_{t=1}^{T} w_t s_t L_t.$0 samples, Top-1 accuracy improves from 91.0% to 94.1%. For the $L_{\text{total}} = \sum_{t=1}^{T} w_t s_t L_t.$1 split, performance rises from 82.8% to 87.5%; for $L_{\text{total}} = \sum_{t=1}^{T} w_t s_t L_t.$2, from 76.0% to 80.8%; and for $L_{\text{total}} = \sum_{t=1}^{T} w_t s_t L_t.$3, from 60.7% to 64.5%. On Rijks’14, Top-1 artist accuracy improves from 63.9% to 67.0% in the 374-artist setting, and from 50.3% to 52.2% in the 374+unknown setting. The paper attributes these gains to correlated attributes acting as “soft priors,” to a shared representation that jointly encodes style, material texture and period cues, and to improved generalization on rare artist classes. A detailed confusion analysis of Jan versus Caspar Luyken is presented as an ablation example in which co-predicted period and material labels reduce errors that a single-task artist network could not resolve. The reported design also cuts training and inference time by 2–3× compared to training four separate networks.

3. The artist embedded in scientific collaboration

A second usage of “Kunster” places the artist inside scientific workflows rather than inside a label space. In the eXtreme19 project, the scientific setting was an international conference on “extremely energetic emission from galaxies,” in particular extreme blazars: Active Galactic Nuclei whose ultra-relativistic jets peak in the $L_{\text{total}} = \sum_{t=1}^{T} w_t s_t L_t.$4-ray band. The conference emphasized multi-wavelength and multi-messenger approaches. The physical framing included the Lorentz factor

$L_{\text{total}} = \sum_{t=1}^{T} w_t s_t L_t.$5

and the Doppler factor

$L_{\text{total}} = \sum_{t=1}^{T} w_t s_t L_t.$6

which govern the apparent brightness and beaming of the jet as seen by an observer (Prandini et al., 2021).

The collaboration was organized through the Italian programme PCTO, “Percorsi per le Competenze Trasversali e per l’Orientamento,” a high-school-to-work programme mandating 100–200 hours of internship-style experience. Prof. Barbara Montolli proposed the project to her 3rd-year students at Liceo Artistico Statale Amedeo Modigliani in Padova, and three underage volunteers—Emily, Beatrice, and Chiara—were selected based on interest, artistic aptitude, and willingness to engage with scientific content. Their 6-month stage had three explicit objectives: immersion in astrophysical concepts behind relativistic jets and spectral energy distributions; development of both a Conference Logo and the main poster image; and learning to balance scientific accuracy and aesthetic appeal through iterative feedback with university researchers.

The process was structured in stages. Joint meetings with Dr. Elisa Prandini, Dr. Michele Doro, and Prof. Daria Mauri introduced SMBHs, jet composition, and the observational roles of Lorentz and Doppler factors. The students reviewed common schematic blazar images and spectral energy-distribution plots featuring the characteristic double-hump, then brainstormed how to encode “multi-wavelength” as coloured stripes and “multi-messenger” as an additional stripe. Preliminary pencil and digital sketches were critiqued for scientific fidelity and compositional balance. Revisions refined the jet anatomy on one side and stylized coloured stripes on the other. A Padova skyline—Sant’Antonio, the Prato della Valle monument, and Porta Ognissanti—anchored the location. In the logo design, the “M” in eXtreme19 was morphed into a two-hump SED, while the “X” was shaped to resemble a relativistic jet piercing through the SED curve.

The outputs were concrete and public. A web-ready logo, poster, and accompanying skyline illustration were used throughout the conference website, email invitations, and on-site banners. Following eXtreme19, the organizers authored a Perspective article in Nature Astronomy, and the editor invited them to propose a cover image. Emily, Beatrice, and Chiara’s poster drawing was selected for Volume 4, Issue 2, February 2020. The achievement was highlighted by Corriere della Sera, Il Sole 24 Ore, and university press offices. The project identifies mutual respect, regular interaction, joint supervision, and six months of iteration—not a single “cold assignment”—as key success factors.

At the South Pole, the artistic role is configured differently but with similar depth. Donald Fortescue entered the scientific environment of the South Pole on an NSF Artists and Writers Fellowship in the austral summer of 2016–17, was shown the IceCube array, and mapped its eighty-six vertical strings of optical sensors onto the eighty-eight notes of a grand piano (Fortescue et al., 2019). He conceived two parallel streams of data: the slow optical sweep of the Sun’s unsetting disk at $L_{\text{total}} = \sum_{t=1}^{T} w_t s_t L_t.$7 S and the barrage of atmospheric muons through the ice. These became the composite artwork Axis Mundi, together with the related sonification 86 Strings #1.

The timelapse component used a weather-sealed DSLR with a wide-angle lens, mounted on a fixed tripod and kept at constant $L_{\text{total}} = \sum_{t=1}^{T} w_t s_t L_t.$8 working temperature. The intervalometer collected one frame every ca. 10–12 seconds over 24 hours, yielding roughly 7 000–9 000 stills, which were compiled at 24 frames per second into a 4 min 48 s HD video. The sonification sampled one muon event per hour, for 24 events total, each chosen from those arriving within $L_{\text{total}} = \sum_{t=1}^{T} w_t s_t L_t.$9 of the instantaneous solar azimuth. Pitch mapping assigned IceCube’s strings 1–86 to the piano’s keys $s_{\text{period}}$0–$s_{\text{period}}$1 in strict sequence. Dynamics mapping linearly mapped Cherenkov–photon hits on each DOM to MIDI-note velocity, and temporal mapping stretched the real-time event duration by approximately $s_{\text{period}}$2 so that each event unfolds over 10–20 s of sound. Custom Python scripts in the IceCube software framework filtered the event files, applied Gaisser-type angular flux models $s_{\text{period}}$3 only to confirm expected rate variations, and emitted a standard MIDI file rendered through a sampled grand-piano virtual instrument.

Both collaborations treat the artist as an informed participant rather than as a decorative adjunct. In Padova, that role produced scientifically robust outreach graphics; at the South Pole, it produced a rule-based audiovisual work in which “sight and sound emerge from the very phenomena under study.” A plausible implication is that, in these cases, Kunster names a mode of co-production in which scientific constraints are part of the artistic method rather than a limit external to it.

4. Artist-centred AI and the redefinition of desirable system properties

“Toward an Artist-Centred AI” proposes a framework of four “atomic” attributes—unexpectedness, diversification, affordance, and evolvability—intended to summarize the values of AI deemed important for artistic creativity (Krekovic et al., 2024). The framework is explicitly positioned against a purely technical ontology of specific models and instead anchors itself in the lived realities of art production, distribution, consumption, and monetization. The motivating claim is that mainstream debates around transparency, fairness, and accountability have limited traction in the artworld, where black-box generative tools may produce eye-catching artefacts yet tend to impose their own aesthetics and biases on artists.

Unexpectedness is defined as the capacity of an AI-powered tool to generate forms, narratives, or processes that exceed conventional boundaries of style, genre, or technique while remaining meaningful or poetically resonant. The paper provides no dedicated mathematical metric, describing it instead through the leveraging of high-dimensional latent spaces and the transcendence of conventional expressive boundaries. The cited examples include DeepDream / Inceptionism, Jake Elwes’s Closed Loop Supercut (2017), and François Quévillon’s Algorithmic Drive (2018–19). The associated design recommendations are conceptually strong prompting, careful curation of outputs, custom training-data pipelines, and hooks for low-level intervention such as layer activations and latent-space walks.

Diversification is distinguished from unexpectedness by its emphasis on continual traversal and fusion of multiple expressive vocabularies, genres, and modalities. It is characterized as “intersective diversification of conceptual, methodological, formal, and aesthetic spaces,” and as a way of avoiding stylistic convergence into predictable or clichéd signatures. Examples include Muse AI Supercuts (Branger_Briz, 2017), Allison Parrish’s Articulations (2018), and broader subgenres such as GANism and deepfakes. The proposed design response is interoperability: import from external libraries, multi-modal interfaces, and open extensions that allow mini-genres to emerge organically.

Affordance concerns accessibility, intelligibility, and control. High-affordance tools offer multi-level configurability, integrate into existing workflows, and make AI “a genuine extension of the artist’s hand rather than an opaque oracle.” The paper again avoids formal metrics and instead enumerates the surrounding ecosystem: pre-trained versus trainable models, web-based GUIs, Jupyter/Colab notebooks, plug-ins for DAWs or 3D suites, no-code AutoML platforms, and code-first SDKs. Hugging Face model hubs, Google Colab notebooks, AWS and Azure AI marketplaces, Microsoft Power Apps’ AI Builder, Apple’s Create ML, and Magenta’s open-source plugins for Ableton Live are given as examples.

Evolvability is defined as the ability of AI-based art tools and practices to grow with rapidly shifting ML research, new data-collection possibilities, and changing cultural contexts. It is described as the “symmetrical adaptability of unexpectedness, diversification, and affordance,” and as sensitivity to institutional and market speeds. Examples include Libby Heaney’s Resurrection (TOTB) (2019), Disnovation.org’s Predictive Art Bot, and Guido Segni’s Demand Full Laziness (2018–23). The engineering recommendations are modularity, microservices or plugin stacks, clear versioning and migration paths across framework shifts such as TensorFlow to PyTorch to JAX, and community-driven extensions.

Across the artistic lifecycle, the framework assigns these attributes specific functions. In creation, unexpectedness and diversification support novelty, affordance democratizes access, and evolvability preserves continuity under technical change. In distribution, diversified outputs and high affordance widen dissemination. In consumption, unexpectedness and diversification sustain engagement while affordance expands participatory roles. In monetization, differentiation resists style fatigue, broader affordance expands collaboration and patronage, and evolvability supports new business models. This framework therefore relocates the question of Kunster from what a system can generate to what conditions preserve human artistic agency within AI-mediated production.

5. Kunster as a mobile neural style-transfer system

“Kunster -- AR Art Video Maker -- Real time video neural style transfer on mobile devices” uses the term as the name of an application implementing real-time neural style transfer on mobile hardware (Dudzik et al., 2020). The mathematical formulation begins with content image $s_{\text{period}}$4, style image $s_{\text{period}}$5, and stylized output $s_{\text{period}}$6, using a frozen VGG-16 pretrained on ImageNet to extract feature maps $s_{\text{period}}$7 at various layers. The content loss is

$s_{\text{period}}$8

the style loss is Gram-matrix based,

$s_{\text{period}}$9

$1/10$0

and the total variation regularizer is

$1/10$1

The stage-1 loss is

$1/10$2

Temporal coherence is introduced through stage-2 fine-tuning on consecutive video frames $1/10$3 from MPI-Sintel, using ground-truth dense optical flow $1/10$4 and occlusion masks $1/10$5. The feature-level temporal loss is

$1/10$6

and the output-level temporal loss is

$1/10$7

The combined two-frame loss adds static losses on both $1/10$8 and $1/10$9, plus $= 0.1$0 and $= 0.1$1. At inference time no flow or masks are needed, because temporal consistency is “baked in.”

The feed-forward network is a variant of ReCoNet’s encoder–decoder plus residual blocks. All convolutions are $= 0.1$2 with reflection padding, all normalization is instance-norm, and Leaky or plain ReLU follows each convolution except the final linear output. The unscaled layer sequence is: input; convolution with 48 filters; convolution with 96 filters and stride 2; convolution with 192 filters and stride 2; four residual blocks at 192 channels; nearest-neighbor upsampling; convolution with 96 filters; nearest-neighbor upsampling; convolution with 48 filters; and a final convolution to 3 channels. Two hyper-parameters modulate size and speed: $= 0.1$3 scales the number of filters, and $= 0.1$4 scales the number of residual blocks.

Training follows a two-stage regime. Stage 1 uses COCO and MPI-Sintel frames resized to $= 0.1$5, with Adam at learning rate approximately $= 0.1$6 and convergence after $= 0.1$7–$= 0.1$8 iterations. Stage 2 uses MPI-Sintel at $= 0.1$9, lowers the learning rate to approximately $w_t$0, and chooses $w_t$1 comparable to the style weight $w_t$2, for example $w_t$3–$w_t$4. The intended result is strong style in stage 1 and flicker-free video refinement in stage 2.

The paper reports representative performance on iPhone 11 Pro and iPhone 6s, averaged over a $w_t$5, 25 FPS test clip with input resolutions of $w_t$6 and $w_t$7. For the full-size model $w_t$8, the network has approximately 470 K parameters and occupies 1.8 MB on disk, achieving 12 FPS and 19.6 FPS on iPhone 11 Pro, and 5 FPS and 9.5 FPS on iPhone 6s. For the medium model $w_t$9, the network has approximately 122 K parameters and occupies 0.47 MB, achieving 21.9 FPS and 25.2 FPS on iPhone 11 Pro, and 13.3 FPS and 19.4 FPS on iPhone 6s. For the small model $224 \times 224$0, the network has approximately 15 K parameters and occupies 0.06 MB, achieving 40 FPS and 54 FPS on iPhone 11 Pro, and 24 FPS and 38 FPS on iPhone 6s. The paper states that speed grows roughly logarithmically as parameters shrink, and that quantizing to INT8 showed no further GPU boost on these phones.

Deployment details matter to the system’s identity. On iOS, CoreML automatically picks NPU versus GPU versus CPU and natively supports the required operations. On Android, the paper reports fragmentation in TensorFlow-Lite and NNAPI backends, unsupported kernel sizes, strides, or padding that may silently fall back to CPU execution, and a 30–40% FPS drop from extra Transpose operations introduced when converting from PyTorch’s NCHW format to TFLite’s NHWC format via ONNX. The stated future directions include more uniform NNAPI support across SoCs, better conversion toolchains, exploration of depthwise-separable or grouped convolutions, and on-device pruning or neural-architecture search targeted at mobile GPU kernels.

6. Cross-cutting themes

Across these works, several recurrent patterns are explicit. OmniArt uses a single shared representation layer with separate task-specific heads; eXtreme19 uses weekly meetings, shared critiques, and joint supervision; the South Pole collaboration adopts a fixed, rule-based procedure rather than composing by ear or eye; the artist-centred AI framework recommends tiered interfaces, documentation of model limitations and data provenance, and interpretability hooks; and the mobile Kunster system uses a two-stage “image then video” training strategy with temporal-coherence fine-tuning (Strezoski et al., 2017, Prandini et al., 2021, Fortescue et al., 2019, Krekovic et al., 2024, Dudzik et al., 2020).

These commonalities suggest that Kunster, across its multiple usages, is consistently associated with structured mediation rather than unstructured spontaneity. In one case, mediation occurs through metadata and loss balancing; in another, through educational supervision and iterative graphical revision; in another, through algorithmically fixed transduction rules; in another, through interface and governance design; and in another, through compression, scaling, and deployment constraints on mobile hardware.

The cited literature also does not support treating Kunster solely as a synonym for generative image production. It includes artist recognition, artist participation in scientific environments, artist-facing evaluation criteria for AI tools, and video stylization systems optimized for real-time mobile inference. Nor does it reduce artistic relevance to benchmark performance alone. The eXtreme19 and South Pole projects foreground co-production and sensory translation; the artist-centred AI framework foregrounds artistic autonomy, authenticity, and livelihood; OmniArt foregrounds semantic entanglement across artist, date, material, and object type; and the mobile application foregrounds usability by non-expert users together with computational efficiency.

Taken together, these lines of work define a field of reference in which Kunster can signify the artist, the recognition of the artist, or the technical systems built for artists. The convergence lies less in a single unified definition than in a shared concern with how artistic practice is represented, supported, and operationalized under contemporary computational and scientific conditions.