Miju Labs

← All buyers · Midjourney

John Joon Young Chung

Research Scientist, Midjourney · Midjourney

high confidenceTier A · Talk to firstImage, video & audio modelstext · image · multimodal

Chung is a research scientist at Midjourney working on AI for creative activity, and the first author of LiteraryTaste — a preference dataset built by having 60 people annotate 100 pairs of creative texts each, then modelling individual versus collective taste. He is, as far as public evidence goes, the person inside a major image lab who has most explicitly treated taste itself as a dataset problem. That makes him the single best-matched researcher in this batch to what Julian sells.

Background

Chung's research sits in HCI and creativity support: interfaces and models for storytelling, illustration and writing. His recent output runs through the top HCI and NLP venues — Toyteller at CHI 2025, Phraselette at DIS 2025, 'LLMs Behind the Scenes: Enabling Narrative Scene Illustration' at EMNLP 2025, and 'Modifying Large Language Model Post-Training for Diverse Creative Writing' at COLM 2025.

He works at Midjourney with a midjourney.com address and describes his focus as aesthetic AI models and new interaction methods for creative work. LiteraryTaste (Nov 2025) is a collaboration spanning Midjourney, NYU, Berkeley, Washington and Notre Dame.

What they run now

Aesthetic and creative AI at Midjourney: how models can be steered toward an individual's taste rather than a population average, and how creative tools should expose that control. LiteraryTaste is the clearest statement of the programme — it separates what people say they like (stated preference) from what they actually pick (revealed preference), finds stated preference has limited predictive value, and shows a finetuned encoder reaching 75.8% accuracy on personal preference versus 67.7% on collective preference. The COLM 2025 post-training paper attacks the complementary problem of creative-output diversity collapse.

What they buy

Evidenced in kind, if not in cash. LiteraryTaste is exactly the artefact a vendor like Miju would sell: recruited human annotators, a paired-comparison protocol, 6,000 judgements, per-annotator identity retained so individual preference can be modelled. The paper's public abstract does not state compensation or recruitment channel, so whether Midjourney paid annotators and through whom is not verified here. The natural next purchases are the same protocol at larger N, in image and video rather than text, with rater provenance and professional qualification attached — and Midjourney already runs a large-scale aesthetic-ranking loop with its own users, which is free but unqualified.

Career

  1. c. 2023 – presentResearch Scientist, Midjourney — Per his own site and midjourney.com email
  2. c. 2018 – c. 2023PhD researcher, HCI / creativity support, University of Michigan — Affiliation inferred from publication record; not confirmed here
  3. c. 2018–2023University of Michigan, PhD, computer science / HCI (unconfirmed)

On the record

taste is individual, not average · 2025-11

Finds people genuinely diverge on creative preference, and that modelling a person's own preference beats modelling the collective. source

self-report is unreliable · 2025-11

Reports that what people say about their tastes has limited utility for predicting what they actually choose. source

research aim · 2026

Describes his work as building AI/ML for creative activity, with an emphasis on aesthetic models and new interaction methods. source

Recent posts

New #CHI2024 paper - we explored the concept of AI bridging authors and audiences in creative language arts for scalable personalization and investigated authors' thoughts on such a media. More details in Taewook's thread!

Interviews & talks

Why it matters here

Top three in this batch. He has already done the thing Julian wants to sell, at small scale, in the wrong modality, inside the right company. The pitch writes itself: LiteraryTaste at 60 participants and text-only is a proof of concept; the image and video version with professionally qualified raters is an operations problem, not a research problem, and that is what a vendor is for. He is also unusually reachable for a frontier-lab researcher — public site, public email, active X.

How to reach

Open with the paper, not the company. Speculation but well-founded: message on X (@john_jyc) or email with a concrete design for LiteraryTaste-for-images — same stated-versus-revealed-preference structure, 100 pairs per rater, raters drawn from named professional populations (art directors, photo editors, colourists) so per-rater expertise is a modelled variable rather than noise. Offer to run a 30-rater pilot at cost. Do not lead with capacity or price; lead with the study design and what it would let him claim that the 60-participant version cannot.

What we could not establish
  • Whether LiteraryTaste's annotators were paid and through which platform or vendor — not stated on the abstract page.
  • His PhD institution and year (inferred, not confirmed).
  • Whether Midjourney has an internal budget line for external annotation or whether this was a research-collaboration arrangement.

Why they are on the map

Midjourney research scientist on 'aesthetic AI models'. Lead author of LiteraryTaste (a paid pairwise taste dataset) and of the diversity-DPO work. source — evidence dated 2025-11

Shapes what 'good data' means.

How Midjourney buys data

Personalization (V7, April 2025) trains on users' own image rankings, which gives Midjourney large first-party preference data. Its research group (Chung, Kreminski et al.) built LiteraryTaste (2025-11) by paying 60 Upwork annotators $100 each for pairwise creative-writing taste labels, and published diversity-aware DPO for creative writing (2025-03). No public contracts with commercial human-data vendors. The careers board (2026) lists only Product, Infrastructure and Medical roles, none for data.

Very lean org run by founder David Holz (~40 to ~160 staff by different 2025 estimates). A small research group works on creative AI and aesthetics (storytelling, taste, personalization). Also has hardware and Midjourney Medical divisions.

Known vendors: Upwork (research annotator recruitment)

Sources

  1. LiteraryTaste: A Preference Dataset for Creative Writing Personalization (arXiv 2511.09310)
  2. John Joon Young Chung — personal site
  3. Google Scholar — John Joon Young Chung