Miju Labs

← All buyers · Google DeepMind

O

Oliver Wang

Principal Scientist, Google DeepMind — leads the Nano Banana team · Google DeepMind

high confidenceTier A · Talk to firstFrontier labsimage

Research lead for Nano Banana and technical lead for Gemini 2.5 Flash Image, with fourteen years at Disney Research Zürich and Adobe Research before Google DeepMind. He is the rare frontier-model research lead who spent a decade building tools for professional creatives, and he has said publicly that human evaluation is central to how they judge the model and that training on AI-generated images is a real risk.

Background

Wang did his PhD at UC Santa Cruz after Cornell, then six years at Disney Research Zürich, then Adobe Research from 2015 to 2023, rising to Principal Scientist. At Adobe he worked on image and video synthesis and editing — the research that feeds Photoshop and Premiere — which means his instincts about what a professional needs from an image tool are formed by a decade of proximity to working creatives rather than by benchmarks.

He moved to Google DeepMind in 2023 and has shipped across the generative-media stack: Imagen 4, Veo 3, Gemini Omni Flash and the Gemini/Nano Banana image family. On the TWIML podcast (September 2025) he framed the central design tension as visual appeal against factual precision — a model can make a prettier picture or a truer one — and flagged the contamination risk of training on synthetic images. He is an area chair at CVPR and NeurIPS.

What they run now

Research leadership of Google DeepMind's image generation and editing models — the Nano Banana / Gemini Image family and adjacent work on Imagen and Veo. The research questions he owns are the ones human data bears on directly: how to trade visual appeal against factual fidelity, how to keep iterative edits coherent, and how to avoid degradation from training on model-generated images.

What they buy

Inference, but well-supported by his stated positions. He needs authentic, human-authored, rights-clean visual data (explicitly not AI-generated), and human preference data that can adjudicate the appeal-versus-accuracy trade-off rather than collapsing it. Concretely: professional photography and design work with provenance; preference judgements that separate 'more attractive' from 'more faithful to the prompt/source'; editing-fidelity judgements across multi-turn sequences. His Adobe background means he will understand a creative-professional supply story immediately.

Career

  1. 2009 – 2015Post-doc, then Research Scientist, Disney Research Zürich
  2. 2015 – 2023Research Scientist → Principal Scientist, Adobe Research
  3. 2023 – presentPrincipal Scientist; leads the Nano Banana team, Google DeepMind — TL for Gemini 2.5 Flash Image; shipped Imagen 4, Veo 3, Gemini Omni Flash
  4. 2006-2010University of California, Santa Cruz, PhD, Computer Science
  5. 2004-2006University of California, Santa Cruz, MS, Computer Science
  6. 1999-2003Cornell University, BA, Computer Science

On the record

Appeal versus accuracy · 2025-09

Building these models means navigating a tension between making images visually appealing and making them factually precise. source

Human evaluation · 2025-09

Human evaluation plays a central role in assessing how the model is performing. source

Synthetic data contamination · 2025-09

Training on AI-generated images rather than authentic sources carries real risk. source

Interviews & talks

Why it matters here

High. He is the research-side twin of Brichtova and arguably the easier technical sell, because his Adobe decade means he does not need to be persuaded that professional creative judgement is a distinct and scarce thing. His stated worry about synthetic-data contamination is a direct argument for buying verified human-authored creative work with provenance — which is a product Julian can define.

How to reach

Academic register. Speculation: he is reachable as a researcher — personal site, GitHub, conference circuit, and as a CVPR/NeurIPS area chair. Open with the provenance problem he raised: a proposal for human-authored, provenance-tracked creative data with the appeal-versus-accuracy trade-off explicitly labelled in the annotation schema. A short preprint or a CVPR-workshop-shaped artefact would reach him better than an email.

What we could not establish
  • X handle appears to be @oliver_wang2 but was not confirmed against a bio naming Google DeepMind; left null.
  • Whether he has any role in data procurement or only in research direction.
  • His current title vs Brichtova's — research and product are separate ladders here.

Why they are on the map

Nano Banana research lead guest on a16z podcast source — evidence dated 2025-10

Needs the data for a model; writes the spec.

How Google DeepMind buys data

Imagen 3 report (2024-08) collected 366,569 ratings from 3,225 raters in 71 countries via an unnamed 'external platform' under GDM data-enrichment policy (living wage). Veo 3.1 page cites human side-by-sides on MovieGenBench (1,003 prompts), VBench I2V and internal sets. Nano Banana leads (2025) describe large-scale human evals plus trusted testers and artist reviews; Gemini Data team job post (2026) covers data curation/generation/evaluation.

Aug 2026: Kavukcuoglu runs GDM incl. Gemini as SVP; Hassabis chair. Gemini has Data, Post-Training, Safety Post-Training and Model Behavior teams (job posts); Gemini app sits inside GDM under Josh Woodward. Gen-media (Imagen/Nano Banana/Veo/Flow) has product lead Brichtova and modality research leads (Wang, Erhan).

Sources

  1. Oliver Wang — personal homepage
  2. Inside Nano Banana and the Future of Vision-Language Models (TWIML #748)
  3. Google DeepMind Developers: How Nano Banana Was Made
  4. Oliver Wang, LinkedIn