Artist in Computational Research

Updated 9 July 2026

Artist is a multifaceted entity that functions as an identity label, supervisory signal, and conditioning variable across music, visual, and audio generative tasks.
Research leverages artist metadata for enhanced recognition, transfer learning, and symbolic sequence modeling, improving metrics like EER and accuracy.
Artist also denotes acronymic systems—from text-to-image diffusion frameworks to Android instrumentation—illustrating its role as both a cultural construct and technical branding tool.

Searching arXiv for the provided topic and cited papers to ground the article in current arXiv records. arXiv search query: ARTIST (Park et al., 2018, Zhang et al., 2024, Backes et al., 2016) artist recognition artist-centred AI artist attribution text-to-audio artist presence In contemporary computational research, artist denotes several distinct but overlapping objects: an identity label for classification and verification, a source of weak or factual supervision, a target of attribution in visual forensics, a stylistic anchor in generative models, and a stakeholder in governance debates around AI systems. The term also appears as an acronymic system name in unrelated technical literatures, notably for text-rich image generation and Android runtime instrumentation. Across these uses, the artist is treated less as a purely aesthetic notion than as a formal entity embedded in representation spaces, optimization objectives, metadata schemas, and regulatory claims (Park et al., 2018).

1. Artist as a recognition target in music information retrieval

Automatic artist recognition models the artist’s musical style by assigning a musical excerpt $x$ either to one of $A$ known artist identities or to a claimed artist identity. The two canonical tasks are identification, with prediction

$\hat y = \arg\max_{i=1\ldots A} S(x,\mu_i),$

and verification, in which a similarity score $S(x,\mu_i)$ supports an accept/reject decision. The corresponding metrics are classification accuracy and Equal Error Rate (EER), the operating point where false-acceptance rate equals false-rejection rate (Park et al., 2018).

A representative hybrid formulation combines a generative i-vector branch with a discriminative deep convolutional neural network branch. In the i-vector branch, 20-dimensional MFCCs are extracted from each frame, a UBM with $C=256$ Gaussians is trained, and a low-rank total variability matrix $T \in \mathbb{R}^{(C\cdot 20)\times R}$ with $R\approx 400$ is learned so that the adapted mean supervector satisfies $M=m+Tw$ , with $w\sim N(0,I)$ . Given sufficient statistics, the posterior mean of the i-vector is

$w = (I + T^T \Sigma^{-1} N T)^{-1} T^T \Sigma^{-1} (F - N m).$

PLDA is then used for scoring. In the DCNN branch, the input is a 3-second mel-spectrogram segment with 128 mel bins and approximately 130 time frames at 23 ms hop; five convolutional layers, followed by a 256-unit fully connected layer and softmax over artists, produce a 256-dimensional deep feature $A$ 0 that is likewise scored with PLDA after averaging over 15 enrollment tracks (Park et al., 2018).

Two fusion regimes are defined. Early fusion concatenates $A$ 1 and applies PLDA. Late fusion averages the independent scores,

$A$ 2

with $A$ 3. On a 500-artist test set with 15 enrollment and 5 test tracks per artist, late fusion yields the lowest artist EER and the highest artist identification accuracy among the compared methods (Park et al., 2018).

Method	Artist EER (%)	Artist Accuracy
i-vector	10.785	0.464
DCNN	8.938	0.408
Early fusion	13.200	0.477
Late fusion	7.813	0.480

The same study reports that the i-vector branch is less sensitive to the number of artist classes and converges with few identities, whereas the DCNN improves steadily as training artists increase. For singer recognition, late fusion reduces EER from 8.257 to 3.241 and raises accuracy from 0.560 to 0.760, reflecting the strong timbral bias of the generative component and the style-dependent capacity of the discriminative component (Park et al., 2018).

2. Artist metadata as supervision, transfer signal, and sequence token

Artist identity has also been used as a factual metadata signal for supervised representation learning. In a Siamese CNN formulation, 3-second audio segments are converted to log-mel spectrograms $A$ 4, and the objective is a multi-negative margin ranking loss over anchor, positive, and $A$ 5 negative samples:

$A$ 6

with cosine distance and margin $A$ 7. Using Million Song Dataset splits of 5,000 artists with 20 songs each, an artist-only model attains 68.0% hold-out artist discrimination accuracy, and frozen artist embeddings outperform genre-trained baselines on GTZAN and FMA-small genre transfer, while joint artist+album+track supervision gives a further small boost on some transfer tasks (Lee et al., 2019).

A related transfer-learning line replaces raw artist IDs with artist group factors (AGFs), defined as low-dimensional latent cluster labels obtained from artist-level bag-of-words representations. MFCC, $A$ 8MFCC, Essentia descriptors, subgenre labels, and main genre are each converted into BoW vectors and then factorized by LDA or clustering, yielding a $A$ 9-dimensional representation $\hat y = \arg\max_{i=1\ldots A} S(x,\mu_i),$ 0 for each artist. DCNNs are trained to predict AGFs and/or genre, penultimate 256-dimensional embeddings are concatenated, and a small MLP predicts the main genre. On FMA, adding all four AGF tasks to genre reduces LogLoss from 0.8891 to 0.7894 in STN and to 0.7727 in MTN, while increasing F1 from 0.5963 to 0.6599 in STN and to 0.6571 in MTN. A “wide STN” with comparable parameter count but no auxiliary tasks performs worse, indicating that the gain is not simply due to capacity (Kim et al., 2018).

Artist identity has further been recast as a sequence vocabulary for playlist title generation. A Transformer encoder-decoder receives artist IDs rather than track IDs as tokens, with artist embeddings looked up from a learned table and decoded autoregressively under standard negative log-likelihood. The motivation is long-tail mitigation: in Melon, over 17% of track IDs appear no more than twice in training, whereas fewer than 5% of artist IDs are that rare; the UNK rate is 14.7% for track IDs versus 3.5% for artist IDs. Under a chronological split that simulates new-release scenarios, artist-ID inputs improve BLEU-1, ROUGE-1, METEOR, BERTScore, SentBERT, and distinct- $\hat y = \arg\max_{i=1\ldots A} S(x,\mu_i),$ 1 diversity on both Melon and MPD. On MPD, for example, BLEU-1 rises from 0.232 to 0.288 and SentBERT from 0.408 to 0.460 (Kim et al., 2023).

Taken together, these results establish artist metadata as a high-yield supervisory source: it is “free” relative to semantic annotation, objective relative to genre, and effective both for metric learning and for symbolic sequence modeling (Lee et al., 2019).

3. Artist attribution and artist-aware style transfer in visual computing

In computational art forensics, artist attribution has been studied as a function of image scale. A large-scale benchmark of 58,630 high-resolution print reproductions by 210 artists from the Rijksmuseum constructed four fixed-pixel-size versions of each image—256×256, 512×512, 1024×1024, and 2048×2048—and trained a fully convolutional AlexNet-style CNN with patch-wise training on random 224×224 crops. Test-time inference fed the entire image through the network once, producing dense class activations that were globally averaged. Mean Class Accuracy rises from 67.0% at 256×256 to 73.4% at 512×512, then plateaus at 73.1% and 72.9% for 1024×1024 and 2048×2048. Artist-wise analysis shows heterogeneity: 119 artists improve with finer scale, 82 decline, and 9 are essentially invariant. Pairwise Pearson correlations of per-artist accuracies also fall sharply across disparate scales, from 0.71 for 256 vs 512 to 0.29 for 256 vs 2048, motivating multi-scale CNN fusion for attribution tasks (Noord et al., 2015).

The artist also appears as a textual style descriptor in arbitrary image style transfer. TxST uses a fixed VGG-19 content encoder and a frozen CLIP ViT-B/32 style encoder, where the style may be an example painting or a text prompt such as an artist name. A positional mapper reconstructs spatial structure from the 512-dimensional CLIP code, and a polynomial cross-attention module fuses normalized content and style features:

$\hat y = \arg\max_{i=1\ldots A} S(x,\mu_i),$ 2

with $\hat y = \arg\max_{i=1\ldots A} S(x,\mu_i),$ 3 in the final model. Training minimizes a weighted combination of directional CLIP loss, contrastive similarity loss, CLIP feature loss, VGG style loss, VGG content loss, and identity loss, with weights $\hat y = \arg\max_{i=1\ldots A} S(x,\mu_i),$ 4 (Liu et al., 2022).

On a 13-artist subset, TxST reports the highest deception rate, approximately 77%, compared with AST’s 66% and CLIPstyler’s 44–47%, while maintaining balanced content and style similarity. The same framework also supports multi-style fusion by concatenating prompts such as “Monet + Picasso + Van Gogh.” This indicates that artist identity can be represented as a manipulable textual conditioning variable rather than only as a class label (Liu et al., 2022).

4. Artist-conditioned regions in generative audio models

In text-to-audio systems, the artist has been formalized as a microlocatable region in a joint text-audio latent space $\hat y = \arg\max_{i=1\ldots A} S(x,\mu_i),$ 5. A prompt $\hat y = \arg\max_{i=1\ldots A} S(x,\mu_i),$ 6 is mapped to a descriptor constellation $\hat y = \arg\max_{i=1\ldots A} S(x,\mu_i),$ 7 with embedding

$\hat y = \arg\max_{i=1\ldots A} S(x,\mu_i),$ 8

and proximity to an artist-reference anchor $\hat y = \arg\max_{i=1\ldots A} S(x,\mu_i),$ 9 is measured by cosine similarity,

$S(x,\mu_i)$ 0

The artist-conditioned region is

$S(x,\mu_i)$ 1

Rather than naming the artist directly, public-taxonomy tags from sources such as RateYourMusic and Discogs are assembled into a descriptor set $S(x,\mu_i)$ 2, whose embedding $S(x,\mu_i)$ 3 serves as a coordinate for microlocating $S(x,\mu_i)$ 4 (Coelho, 21 Nov 2025).

The corresponding audit protocol generates $S(x,\mu_i)$ 5 samples conditioned on $S(x,\mu_i)$ 6, embeds each generated audio sample, computes similarities to $S(x,\mu_i)$ 7, and returns the similarity statistics and hit rate

$S(x,\mu_i)$ 8

Using $S(x,\mu_i)$ 9 generations per prompt constellation and $C=256$ 0, the study reports mean similarity $C=256$ 1, standard deviation $C=256$ 2, and hit rate for four case studies: Bon Iver (0.82, 0.08, 40%), Philip Glass (0.77, 0.06, 50%), William Basinski (0.74, 0.05, 30%), and Panda Bear (0.79, 0.07, 35%). Stability analyses further report that once an output enters $C=256$ 3, remix/regenerate operations preserve core features with mean shift $C=256$ 4 over successive iterations (Coelho, 21 Nov 2025).

The paper characterizes these proximities as sonic fingerprinting, compositional resonance, aesthetic aura, and hybrid (vocal+avant-pop), and argues that interface-level filters blocking explicit names are easily circumvented by reordering tags. This suggests that artist identity in generative audio is not merely a semantic label but a navigable and reproducible latent locus (Coelho, 21 Nov 2025).

5. The artist as stakeholder in AI design, transparency, and ownership

A distinct research line treats the artist as the normative center of AI system design. An artist-centred AI framework proposes four atomic attributes of “artistically desirable AI”: Unexpectedness, Diversification, Affordance, and Evolvability. Unexpectedness is informally the capacity to help the artist escape habitual forms, with a semi-formal proxy

$C=256$ 5

Diversification may be formalized by Shannon entropy or feature-space variance. Affordance is summarized as $C=256$ 6, where the terms count exposed hyperparameters, data-insertion points, and types of real-time feedback. Evolvability is proxied by

$C=256$ 7

The paper sketches the overall contribution of AI to artistic creativity as $C=256$ 8, with a possible linearization $C=256$ 9 (Krekovic et al., 2024).

The framework is motivated by pressures on art production, distribution, consumption, and monetization: black-box style engines flatten individual expression, recommender systems drive monocultures, and NFT hype enforces faux-scarcity. Concrete examples include DeepDream, DALL·E 2 with negative prompts, Muse AI Supercuts, Allison Parrish’s Articulations, Hugging Face model cards and parameter sliders, Stable Diffusion Web UIs, RunwayML, Custom Vision, LoRA adapters, SAM, and “Predictive Art Bot.” The authors also enumerate trade-offs: unexpectedness versus usability, diversification versus overwhelm, affordance versus complexity, and evolvability versus stability (Krekovic et al., 2024).

Survey evidence complements this design perspective. A mixed-method survey of 459 artists reports that 80.2% agree or strongly agree that AI model creators must disclose in detail what art or images were used to train their systems; 61.9% agree or strongly agree that Gen-AI models are a threat to art workers; and 44.9% agree or strongly agree that Gen-AI art is a positive development. When asked about outputs generated in an artist’s recognizable style, 41.4% assign ownership to the original artist, 39.2% to the user, and 26.8% to the model creator. With respect to compensation, 50.97% report that they do not need direct profit but care who benefits, while 22.8% select “I don’t need profit, but I don’t want for-profit companies to profit from my art” (Lovato et al., 2024).

Ordinal logistic regressions in the same study indicate that artists under 30 are more likely to see AI as a threat (OR≈1.5), artists who have used AI themselves are less likely to view it as a threat (OR≈0.57) and more likely to accept user-ownership (OR≈2.0), and persons of color more strongly favor detailed disclosure (OR≈2.6) and ownership by original artists (OR≈1.95). These results align with recommendations for mandatory training-data disclosure, opt-in licensing and consent tools, shared ownership models, compensation mechanisms, inclusive governance, and ongoing monitoring (Lovato et al., 2024).

6. ARTIST and ARTist as acronymic systems

The term ARTIST also designates a diffusion-based image-generation framework for rendering legible text in images. This system uses a disentangled architecture with a pretrained textual diffusion model and a visual diffusion model initialized from Stable Diffusion v1.4. The text module is trained on synthetic black-and-white text renderings under the noise-prediction loss

$T \in \mathbb{R}^{(C\cdot 20)\times R}$ 0

first on approximately 10M word-level samples for 400K steps and then on approximately 50M sentence-level samples for 200K steps. The visual module receives intermediate decoder features from the frozen text module and is optimized with

$T \in \mathbb{R}^{(C\cdot 20)\times R}$ 1

for 250K steps on MARIO-10M. A pretrained LLM such as GPT-4 is used to extract words, layout, and scene description from the user prompt (Zhang et al., 2024).

On the 5,414-prompt MARIO-Eval benchmark with four images per prompt, TextDiffuser yields OCR-Acc 0.5712, FID 38.76, and CLIPScore 0.3436; ARTIST-TD yields OCR-Acc 0.6298, FID 36.58, and CLIPScore 0.3466; and ARTIST with GPT-4 yields OCR-Acc 0.7373, FID 38.43, and CLIPScore 0.3482. Decoder-feature injection outperforms a ControlNet-style encoder injection, 0.6298 versus 0.4935 in OCR-Acc, and GPT-4 raises keyword-F1 from 0.64 to 0.98 (Zhang et al., 2024).

By contrast, ARTist denotes the Android Runtime Instrumentation and Security Toolkit, a compiler-based instrumentation framework for Android built on ART and dex2oat’s Optimizing backend. ARTist exposes an HOptimization plugin interface, scans each method’s SSA-based HGraph, and inserts, replaces, or deletes HInstructions to realize modules such as an inline reference monitor or performance tracing. Because the APK on disk remains unchanged and only the generated oat file is modified, signature preservation is maintained. In a taint-tracking case study, source and sink instrumentation supports propagation rules such as $T \in \mathbb{R}^{(C\cdot 20)\times R}$ 2 for a binary operation and uses a per-thread stack for local sink/source matching (Backes et al., 2016).

Evaluation reports that ARTist successfully instruments 83 of 85 top non-multidex apps, or 97.6%, with a simple tracing module. Microbenchmarks on a Nexus 5 show overheads of approximately 8.9% for getConfiguredNetworks(), 1.2% for isWifiEnabled(), 30.6% for getScanResults(), and 3.3% for startDiscovery(). For taint tracking, Passmark Java-only benchmarks indicate overheads of approximately 30.7% on CPU, 16.5% on disk, and 7.7% on memory, while a subset of DroidBench yields overall recall of approximately 54/66 = 81.8% (Backes et al., 2016).

These acronymic usages are unrelated in domain, but they illustrate a recurrent naming pattern in which “ARTIST” signals modularity, interpretability, or tool support rather than the human creator. A plausible implication is that the research literature now uses the term at two levels simultaneously: as a modeled cultural identity and as a branding device for technical systems.