Miju Labs

← All buyers · Krea

S

Sangwu Lee

ML Researcher (lead author, FLUX.1 Krea and Krea 2) · Krea

high confidenceTier A · Talk to firstImage, video & audio modelsimage

Lee is the lead author of both the FLUX.1 Krea post-training write-up and the Krea 2 technical report, which together are the most explicit public account anywhere of how a lab builds an opinionated aesthetic into a model. He personally documented rejecting open-source preference datasets because they bias toward symmetric, simple composition and the 'AI look', and building instead on in-house preference labels from people who know the model's specific weaknesses. He is the most articulate taste-engineer in this batch — and, for that reason, the hardest sell.

Background

Lee is an ML researcher/engineer at Krea, working with Erwann Millon on the model side. The FLUX.1 Krea release (July 2025) was a collaboration with Black Forest Labs, which supplied a raw flux-dev base model that Krea post-trained — a direct commercial link between two companies in this batch.

His public talk 'Training Krea 2: What matters in generative model training' and the Krea 2 technical report lay out a full pipeline: captioning, progressive-resolution pretraining, midtraining with FAISS clustering and PageRank-weighted entity coverage, hand-curated SFT, two-stage preference optimisation, multi-reward GRPO-style RL, and TDM distillation.

What they run now

Krea's model training end to end. The post-training detail is the relevant part: preference optimisation runs in two stages — synthetic preference pairs built with an on-policy delta-learning strategy, then calibration against in-house human annotations from people familiar with the model's specific strengths, weaknesses and quirks. RL uses four separate reward models: a general aesthetic reward (a pointwise Bradley-Terry formulation, chosen over pairwise VLM judges for efficiency), a prompt-following reward driven by per-prompt rubrics that decompose intent into verifiable requirements, a text-rendering reward, and an artifact/structure reward that catches extra fingers, malformed limbs and distorted text.

What they buy

Evidenced, and evidenced as a refusal. Krea 2 states that human annotations for preference optimisation were collected in-house rather than through external vendors, and the FLUX.1 Krea post explicitly rejects public preference datasets. So the honest read is: Lee is a sophisticated buyer who has already decided that generic outsourced preference data is worse than nothing. What he might buy is the narrow thing he cannot staff: raters with deep domain expertise Krea does not have internally (typography and text rendering, product and fashion photography, specific illustration traditions), and rubric-driven prompt-following judgements at volumes an in-house team cannot reach.

Career

  1. c. 2024 – presentML Researcher / Machine Learning Engineer, Krea — Lead author, FLUX.1 Krea and Krea 2

On the record

against generic preference data · 2025-07

Says training on open-source preference datasets biased the model toward symmetric, simple composition and regressed it toward the AI look. source

be opinionated · 2025-07

Argues it is often better and easier to overfit a model toward a particular style than to chase a marginal preference distribution that satisfies nobody. source

quality over quantity · 2025-07

Reports needing a surprisingly small amount of data — under a million samples — for effective post-training, with labellers who understood the model's weaknesses. source

SFT data · 2026

States that for supervised finetuning the quality of the dataset matters far more than its scale. source

Interviews & talks

Why it matters here

Top three in this batch, but as the most important objection rather than the easiest sale. Lee has written down, publicly and precisely, why the standard human-preference-data product fails — marginal preference distributions that are not biased enough to satisfy anyone. Julian's whole thesis has to survive that argument, and if it does, Krea's write-ups become the best possible sales material for every other buyer. Read both documents in full before pitching anyone in this market.

How to reach

Speculation: the only credible approach is to concede his point and sell the opposite of what he rejected — not a pooled preference dataset, but a panel whose bias is specified, documented and selectable, so the customer chooses whose taste the model overfits to. Frame it as 'you said marginal preference distributions are useless; here is how to buy a non-marginal one'. Reach him through the Krea 2 GitHub repo or via Erwann Millon, who has public accounts. Do not pitch scale.

What we could not establish
  • No confirmed LinkedIn or X handle found.
  • Who Krea's in-house labellers are — employees, contractors or community members.
  • Whether the BFL relationship means Krea and BFL share any data supplier.

Why they are on the map

First author of the Krea 2 report and co-author of the FLUX.1 Krea post-training write-up on opinionated preference data. source — evidence dated 2026-06

Needs the data for a model; writes the spec.

How Krea buys data

Krea 2 tech report (2026-06-23) says preference annotations were 'collected entirely in house, by people familiar with the specific strengths, weaknesses, and quirks of the model', and that they prototyped reward models with Thinking Machines' Tinker. FLUX.1 Krea post (2025-07) argues that generic preference datasets hurt aesthetics and that under 1M high-quality examples is enough. May 2026: a Contra Labs study using working creatives benchmarked Krea 2 Large on style fidelity. 2026 'ML Researcher - Posttraining' role will 'coordinate with data teams and partners to manage collection of preference data' and 'train custom VLM as reward models'.

Small team (~17 in April 2025, grown since). Founders lead. Research covers Diffusion and Posttraining (RL, reward VLMs, evals). There is an internal data team that does annotation, and a 'Personalization & Recommendation' ML role models user taste.

Known vendors: contra-labs, Thinking Machines Tinker (infra, not data)

Sources

  1. Krea 2 Technical Report
  2. Releasing Open Weights for FLUX.1 Krea (Sangwu Lee & Erwann Millon, 2025-07-31)
  3. krea-ai/krea-2 on GitHub