Beyond an internal judgment layer for Ivan: could historical approvals, Pinterest references, typography preferences, and hundreds of written critiques train a system that understands what feels aligned, generic, emotionally right — and why?
Is the “why” layer achievable? Should comments be embedded, structured into dimensions, taxonomized, or linked to visuals as multimodal supervision — or does mixing thousands of subjective comments become incoherent “gray soup”?
Where is the boundary between visual preference modeling, aesthetic reasoning, and tacit creative judgment? Has anyone explored codified creative-director cognition publicly? Could this become thought leadership — judgment modeling and taste infrastructure, not merely generative AI — with real audience traction?
You’ve now asked the version of the question that’s actually worth answering — and the shift from “could we build this?” to “is this an idea worth being known for?” is the more interesting move. They have different answers, and the second one is more interesting than the first.
The research subfield already exists: explainable image aesthetics assessment or aesthetic critique generation — a continuous thread since 2017. Photo Critique Captioning datasets, Neural Aesthetic Image Reviewers on hundreds of thousands of comments, AesExpert with 88K human feedbacks and instruction-tuning datasets, CG-IAA fine-tuning CLIP on multimodal critique databases with attribute-specific experts.
The path from images plus comments to a model that can score and explain is a paved road. You wouldn’t be inventing the science; you’d be inventing the application to one specific person’s judgment.
Public work uses populations of judges and broad dimensions (“composition,” “lighting”). Your project is the opposite: one judge, idiosyncratic dimensions (“luxury restraint,” “emotional pacing,” “good tension”). The research provides the machinery; you provide the vocabulary — and that vocabulary comes from your existing comments, not from a labeling exercise.
Don’t pre-decide dimensions. Don’t write a “framework of Arseni’s taste” from the armchair. Take 1,000–3,000 actual approvals and rejections with comments, embed the comments, cluster them, let dimensions emerge.
Surface forms like “too generic,” “luxury restraint,” “breaks emotional pacing” are likely expressions of fewer underlying axes: restraint vs. excess, tension vs. resolution, specificity vs. genericness, cinematic vs. literal, atmosphere vs. content.
This step alone — even without any model on top — would be a meaningful intellectual artifact: the first structured map of your taste rather than accumulated instinct.
The “gray soup” risk is real if comments are embedded in isolation. The fix: use comments as labels for visual examples. Each (image, comment) pair becomes training data for a CLIP-style model fine-tuned on your lexicon. Image grounded; comment as attribute label. Gray soup dissolves when comments are supervision anchored to images.
Stack an LLM on top whose job is to compose axis scores into critique prose: “Low on restraint, high on busyness, typography reads generic-luxury rather than specific-luxury — likely rejection, primarily for visual noise.” Recognizably yours because the axes and vocabulary are yours. The model isn’t pretending to be Arseni. It’s reading the image through Arseni’s grammar.
The realistic product: a critique partner capturing the stable 70–80% of judgment, in your vocabulary, with calibrated abstention on the contextual 20–30%. Same shape as the operational layer. Same ceiling. Same value.
Could this become a recognizable intellectual position — not just a tool, but a stance? Yes, with a specific framing, and a window currently open but not permanent.
The mainstream creative-AI conversation (2025–2026) has converged on: taste is the new bottleneck. Taste as learnable leadership capability. As curation and rejection in a world of infinite generation. Craft-as-luxury. Luxury executives moving GenAI from curiosity to execution.
Almost nobody is talking about the synthesis you’re proposing: taste as something that can itself be infrastructure. The conversation is stuck on a binary — AI generates, taste judges, kept separate — or vague “co-pilot” gestures. Your move: taste is not the opposite of AI; taste is what we should be modeling with AI, and studios that operationalize their creative directors’ judgment gain a structural advantage pure-generation studios cannot match.
Where it would not resonate yet: mass design community, generic AI-adoption consulting, agency thought-leadership circuit. They’re on the previous beat. Window: probably 12–18 months.
One: thinking visibly downstream of real practice. You have fifteen years of recorded approvals and rejections from a major artist. Almost nobody writing about AI and taste has a corpus. Make the framework empirically derived, not theoretically derived from other essays.
Two: take a side. The strongest version:
The next decade of elite creative work will not be defined by which studios use generative AI most aggressively. It will be defined by which studios externalized their creative directors’ judgment into infrastructure that survives the director’s attention, scales their consistency, and frees them for work that requires actual presence. Those studios will look more human and more idiosyncratic, not less, because their tools encode specific taste rather than generic competence.
Back it with a working system at iHouseDesign with Drebin as the case study. That becomes a position with your name on it.
The trap: don’t lead with technical architecture. Lead with problem and stance. Ivan and the operational layer are evidence for the stance, not the stance itself.
Technically: the “why” layer is achievable — research exists, path from your corpus is well-mapped. Gray-soup fix: multimodal supervision against axes extracted from your comments.
Strategically: a real position exists; “taste as modelable infrastructure, evidenced by luxury creative ops practice” is currently unoccupied. Cleanest way to occupy it: build the Ivan system as proof, write the position publicly with Drebin as implicit credential, let “taste infrastructure” point back to you.
Coda: building forces articulation; articulation compounds whether or not the model ships. The position exists the moment you start writing about what you discovered while trying to model your own judgment. That essay may be more valuable than the model.