---
schema: 1
type: video
title: >-
  Training Krea 2: What matters in generative model training — Sangwu Lee,
  Krea.ai
description: >-
  The most reliable way to render a person is to render the most boring average
  person and put them in the center of the frame. Sangwu Lee offers that as the
  price the big image models pay for consistency: ask a production model for a
  burning skull and every output comes back clean, competent, and nearly
  identical. Krea 2, whose medium variant is now open source, trades the other
  way, optimizing for fast generation and stylistic range so that a studio that
  does not yet know what it wants can actually explore. Most of the talk is
  about data, which he says twice over is basically everything once the
  architecture is locked. The examples are specific. A painting photographed on
  a wall is perfectly good training data except that captioners consistently
  omit the frame and the white wall behind it, so the model learns to hang every
  painting it generates. They refuse to train on AI generated images at all,
  because the aesthetic is sticky and you inherit somebody else's model.
  Deduplication runs on hashes first across two to ten billion images, then on
  embeddings for near duplicates. A large vision language model's judgment gets
  distilled down into a classifier cheap enough to sweep a billion images.
  Sparse autoencoders double as an unsupervised tagging system for catching
  watermarks and border artifacts. World knowledge coverage is checked against
  Wikipedia concepts ranked by PageRank. Thirty to forty in house filters in
  total. Speaker info: - https://github.com/RE-N-Y -
  https://re-n-y.github.io/devlog/ - https://github.com/krea-ai/krea-2
  Timestamps: 0:00 - Open sourcing Krea 2 medium 1:40 - Consistency versus
  diversity in production models 3:23 - How diffusion models train, and why
  latent space 5:59 - Data is basically everything 6:53 - Bad data, and why they
  refuse AI images 8:34 - The captioning pipeline, and the painting on a white
  wall 10:15 - Deduplication and cheap classifiers at billion image scale 11:56
  - Sparse autoencoders as an unsupervised tagging system 13:39 - Wikipedia
  PageRank for world knowledge coverage 14:35 - The training pipeline, borrowed
  wholesale from LLMs 18:54 - What actually mattered for iterating fast 19:46 -
  The stack is inverting back toward DALL-E 2
canonical: >-
  https://liquidgold.cc/channel/aidotengineer/explore/video/training-krea-2-what-matters-in-generative-model-tr-1f0432b8ac69
markdown: >-
  https://liquidgold.cc/channel/aidotengineer/video/training-krea-2-what-matters-in-generative-model-tr-1f0432b8ac69.md
llms: null
parent: 'https://liquidgold.cc/channel/aidotengineer.md'
channel: aidotengineer
updated: '2026-10-24T05:18:52.979Z'
language: null
children: []
videoId: '-tviRdpmHvs'
durationSeconds: 1306
publishedAt: '2026-08-18T14:00:06.000Z'
searchChunks:
  - ordinal: 0
    text: null
    startSeconds: 23
    endSeconds: 100
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=23'
    heldReason: rights
  - ordinal: 1
    text: null
    startSeconds: 110
    endSeconds: 207
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=110'
    heldReason: rights
  - ordinal: 2
    text: null
    startSeconds: 110
    endSeconds: 207
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=110'
    heldReason: rights
  - ordinal: 3
    text: null
    startSeconds: 219
    endSeconds: 364
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=219'
    heldReason: rights
  - ordinal: 4
    text: null
    startSeconds: 219
    endSeconds: 364
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=219'
    heldReason: rights
  - ordinal: 5
    text: null
    startSeconds: 374
    endSeconds: 421
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=374'
    heldReason: rights
  - ordinal: 6
    text: null
    startSeconds: 421
    endSeconds: 431
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=421'
    heldReason: rights
  - ordinal: 7
    text: null
    startSeconds: 431
    endSeconds: 524
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=431'
    heldReason: rights
  - ordinal: 8
    text: null
    startSeconds: 431
    endSeconds: 524
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=431'
    heldReason: rights
  - ordinal: 9
    text: null
    startSeconds: 536
    endSeconds: 615
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=536'
    heldReason: rights
  - ordinal: 10
    text: null
    startSeconds: 615
    endSeconds: 626
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=615'
    heldReason: rights
  - ordinal: 11
    text: null
    startSeconds: 626
    endSeconds: 719
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=626'
    heldReason: rights
  - ordinal: 12
    text: null
    startSeconds: 730
    endSeconds: 824
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=730'
    heldReason: rights
  - ordinal: 13
    text: null
    startSeconds: 824
    endSeconds: 835
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=824'
    heldReason: rights
  - ordinal: 14
    text: null
    startSeconds: 895
    endSeconds: 1134
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=895'
    heldReason: rights
  - ordinal: 15
    text: null
    startSeconds: 895
    endSeconds: 1134
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=895'
    heldReason: rights
  - ordinal: 16
    text: null
    startSeconds: 895
    endSeconds: 1134
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=895'
    heldReason: rights
  - ordinal: 17
    text: null
    startSeconds: 895
    endSeconds: 1134
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=895'
    heldReason: rights
  - ordinal: 18
    text: null
    startSeconds: 895
    endSeconds: 1134
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=895'
    heldReason: rights
  - ordinal: 19
    text: null
    startSeconds: 1202
    endSeconds: 1306
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=1202'
    heldReason: rights
  - ordinal: 20
    text: null
    startSeconds: 1202
    endSeconds: 1306
    watch: 'https://liquidgold.cc/watch/-tviRdpmHvs?t=1202'
    heldReason: rights
---


# Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai

The most reliable way to render a person is to render the most boring average person and put them in the center of the frame. Sangwu Lee offers that as the price the big image models pay for consistency: ask a production model for a burning skull and every output comes back clean, competent, and nearly identical. Krea 2, whose medium variant is now open source, trades the other way, optimizing for fast generation and stylistic range so that a studio that does not yet know what it wants can actually explore. Most of the talk is about data, which he says twice over is basically everything once the architecture is locked. The examples are specific. A painting photographed on a wall is perfectly good training data except that captioners consistently omit the frame and the white wall behind it, so the model learns to hang every painting it generates. They refuse to train on AI generated images at all, because the aesthetic is sticky and you inherit somebody else's model. Deduplication runs on hashes first across two to ten billion images, then on embeddings for near duplicates. A large vision language model's judgment gets distilled down into a classifier cheap enough to sweep a billion images. Sparse autoencoders double as an unsupervised tagging system for catching watermarks and border artifacts. World knowledge coverage is checked against Wikipedia concepts ranked by PageRank. Thirty to forty in house filters in total. Speaker info: - https://github.com/RE-N-Y - https://re-n-y.github.io/devlog/ - https://github.com/krea-ai/krea-2 Timestamps: 0:00 - Open sourcing Krea 2 medium 1:40 - Consistency versus diversity in production models 3:23 - How diffusion models train, and why latent space 5:59 - Data is basically everything 6:53 - Bad data, and why they refuse AI images 8:34 - The captioning pipeline, and the painting on a white wall 10:15 - Deduplication and cheap classifiers at billion image scale 11:56 - Sparse autoencoders as an unsupervised tagging system 13:39 - Wikipedia PageRank for world knowledge coverage 14:35 - The training pipeline, borrowed wholesale from LLMs 18:54 - What actually mattered for iterating fast 19:46 - The stack is inverting back toward DALL-E 2

Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/training-krea-2-what-matters-in-generative-model-tr-1f0432b8ac69 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md)

> Search chunk text is held until public-source provenance is verified.

## Document chunks

> Missing: no stored document chunks are available.