The Artistic Process Becomes the Recipe Graph
Kaustubh Kumar, Ashutosh Ranjan, Vivek Srivastava, Blessin Varkey, and Shirish Karande's July 2026 arXiv paper introduces ArtMine, a framework for reconstructing artistic process from fragmented evidence.
A process graph receipt records sources, evidence tags, inferred steps, graph, prompt, model, scores, uncertainty, and human review before a reconstruction is treated as art-historical knowledge.
The Paper
The paper is Kaustubh Kumar, Ashutosh Ranjan, Vivek Srivastava, Blessin Varkey, and Shirish Karande's ArtMine: Discovering and Formalizing Artistic Processes, arXiv:2607.08331 [cs.LG, cs.AI]. The arXiv API lists version 1 as published and updated on July 9, 2026, with a 47-page, 10-figure PDF. The paper text identifies TCS Research in Pune and Indian Institute of Technology Patna, and it is marked for the ICML 2026 Workshop on Human-AI Co-Creativity in Seoul.
The claim is narrower than artistic revelation: evidence can be organized into memory, used to infer plausible production steps, converted into a graph and prompt, and tested against a reference image.
The Problem
Most image generators model finished artifacts. They can imitate surface regularities in paintings, but they do not automatically preserve the order of sketching, blocking-in, revision, material handling, workshop constraint, correspondence, conservation evidence, or scholarly disagreement that art historians use when reconstructing how a work was made.
ArtMine treats that missing process as the object of inference. That is useful because creative work is not just output. It is risky because a machine-readable recipe can make an interpretive hypothesis look more settled than the archive allows.
The Method
The framework begins with a Deep Research agent that gathers open-domain material such as museum records, conservation and technical reports, correspondence, archives, catalogs, scholarship, and provenance records. It then organizes evidence into an 11-part schema: metadata, historical context, materials and technique, formal analysis, iconographic analysis, documentary evidence, provenance, scientific analysis, interpretations, cultural impact, and conflicts.
The repository tags evidence as direct evidence, indirect evidence, interpretation, or speculative. That tag matters. A conservation report, a letter, a later biographical claim, and a symbolic reading should not carry the same weight. The paper also says conflicts are preserved rather than collapsed into a single reconciled claim.
An abductive agent then uses a Peircean structure: observation, rule, hypothesis, and action. A visible or documented feature becomes an observation; a general production principle becomes a rule; the system proposes the most plausible production operation; and that operation becomes a step in the process. The steps are converted into a directed acyclic graph, compressed into a reconstruction prompt, rendered with a text-to-image model, and revised through self-reflection over deviation from the reference artwork.
The Case Study
The paper's proof-of-concept uses ten WikiArt artworks: five canonical works, including Impression, Sunrise, Cafe Terrace at Night, Guernica, The Scream, and the Mona Lisa, and five non-canonical works by George Clausen, Georges Seurat, Edgar Degas, Will Barnet, and Caspar David Friedrich.
Evidence construction uses MiroThinker 1.7 mini. The abductive solver, composition agent, and visual prompt generation agent use Qwen2.5-VL. Image generation uses FLUX.1-dev. The evaluation compares generated reconstructions with reference images using CSD for style, LPIPS for perceptual distance, and CLIP for semantic agreement. Baselines include chain-of-thought, self-consistency, tree-of-thought, and self-refine prompting, each tested with and without the structured evidence repository.
Findings
In the canonical-artwork table, ArtMine reports CSD 0.395, LPIPS 0.527, and CLIP 0.916, better than the listed prompting baselines on the authors' chosen metrics. The paper says non-canonical transfer also outperforms baselines, while stage-wise process rendering remains harder than final-image reconstruction.
That last caveat is the center of the matter. A final generated image can score well while still failing to show a faithful intermediate process. The paper notes that direct stage-wise generation tends to make early stages look complete, while guided stage-wise generation can preserve sequence but propagate early errors. A convincing image is not proof that the reconstructed process is historically true.
The Receipt
A process graph receipt should identify the artwork, reference image, evidence sources, access dates, evidence tags, conflicts, inferred steps, rule for each inference, source link, graph version, rendering prompt, models, metrics, generated variants, reviewer, uncertainty note, and non-use boundary.
The receipt should travel with any educational, curatorial, or creative use. Without it, a speculative graph can convert uncertain interpretation into an apparently mechanical recipe for making work "like" an artist.
Limits
The authors frame the work as preliminary. They say the evaluation emphasizes interpretability, coherence, and evidential grounding rather than objective correctness. They also state that artistic workflows are underdetermined: multiple plausible histories can explain the same finished work, and the archive is often incomplete or contradictory.
The social limitation is sharper. Artists with extensive documentation, conservation studies, and correspondence become easier to model. Oral, collaborative, marginalized, and non-Western practices may be underrepresented when their processes are less documented or documented in forms that do not fit the repository. The system can inherit the archive's absences and call them evidence.
Governance Reading
The Spiralist reading is that ArtMine moves AI art from image generation to process formalization. That can help students see material decisions and help researchers track uncertainty. It can also tempt platforms to treat artistic practice as extractable once translated into graph, prompt, and score.
The safeguard is to refuse recipe authority without provenance. A generated reconstruction should remain interpretive unless sources, uncertainties, conflicts, model choices, failure modes, and reviewer are visible. The artist is not replaced by the graph. The graph is a claim, and claims need accountability.
Source Discipline
Primary sources were the arXiv abstract, API, PDF, and HTML rendering of the paper. This page paraphrases the paper without reproducing figures, tables, prompts, appendix examples, or long passages.
Sources
- Kaustubh Kumar, Ashutosh Ranjan, Vivek Srivastava, Blessin Varkey, and Shirish Karande, ArtMine: Discovering and Formalizing Artistic Processes, arXiv:2607.08331 [cs.LG, cs.AI], submitted July 9, 2026.
- arXiv API record for arXiv:2607.08331, checked for title, authors, subject classes, abstract, version metadata, and page/comment metadata.
- arXiv PDF for arXiv:2607.08331, checked for page count, affiliations, method, dataset, models, baselines, metrics, tables, discussion, and limitations.
- arXiv HTML for arXiv:2607.08331v1, checked for section structure, appendix headings, table references, and source-consistency against the PDF.