--- title: Training Krea 2: What matters in generative model... description: A Krea.ai researcher describes the training pipeline behind Krea 2, the company's recently open-sourced image foundation model, focusing on stylistic diversity... canonical: https://daily.dev/posts/training-krea-2-what-matters-in-generative-model-training-sangwu-lee-krea-ai-ocrqtpdvc twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai | daily.dev og:description: A Krea.ai researcher describes the training pipeline behind Krea 2, the company's recently open-sourced image foundation model, focusing on stylistic diversity... og:url: https://daily.dev/posts/training-krea-2-what-matters-in-generative-model-training-sangwu-lee-krea-ai-ocrqtpdvc og:image: https://api.daily.dev/og/posts/ocRqTpdVc.png og:image:alt: Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 21 min read · 0 upvotes · 0 comments ## Summary A Krea.ai researcher describes the training pipeline behind Krea 2, the company's recently open-sourced image foundation model, focusing on stylistic diversity as a differentiator from production models like ChatGPT/Nano Banana Pro that tend toward mode collapse for reliability. The talk covers data curation (deduplication via hashing and embedding methods, removing AI-generated synthetic data, using sparse autoencoders as unsupervised tagging systems, Wikipedia page-rank-driven concept coverage), a detailed captioning pipeline using OCR plus vision-language models, and a Baidu/LLM-inspired training progression from low-to-high resolution pretraining through mid-training, SFT, preference optimization, GRPO-style reinforcement learning, and a prompt-expansion model. It closes with speculation on future directions like multi-expert distillation and richer image-conditioning representations (bounding boxes, scene graphs). ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How does Krea 2's approach to image generation differ from ChatGPT and Nano Banana Pro? Krea 2 prioritizes stylistic diversity and fast generation over the slow, highly consistent but mode-collapsed outputs typical of production models like ChatGPT and Nano Banana Pro. Those models tend to render the most generic, centered version of a subject to guarantee reliability, taking up to a minute or two per image, whereas Krea 2 aims for quick iteration across varied visual styles for creative exploration. _daily.dev surfaces training approaches like this for developers comparing image generation models._ ### How do you filter AI-generated images out of a training dataset for an image model? Krea's team trained a large vision-language model to classify whether an image looks AI-generated, then distilled that classifier down into a small, cheap SigLip-sized model so it could run efficiently across billions of images. They avoided training on synthetic data entirely because it makes the resulting model imitate the aesthetic of models like ChatGPT or Nano Banana Pro rather than learning from real data. _Developers curating training data for generative models can track techniques like this on daily.dev._ ### What training pipeline stages are used for modern production diffusion models? Production-grade diffusion models follow a pipeline inspired by LLM training: low-to-high resolution pretraining (from 256px up to 1K), mid-training on curated domain data, supervised fine-tuning, preference optimization using pairwise comparisons, GRPO-inspired reinforcement learning against reward servers, and training a separate small LLM to expand short user prompts into long, detailed prompts before feeding the diffusion model. _daily.dev helps engineers building generative pipelines keep up with evolving diffusion training stages._ ## Similar posts on daily.dev - [The image model race just split into four lanes](https://daily.dev/posts/the-image-model-race-just-split-into-four-lanes-bkf01m54q) · Visually AI · 0 upvotes · 0 comments - [The state of open source AI models in 2025](https://daily.dev/posts/the-state-of-open-source-ai-models-in-2025-nbwk4nxtr) · Red Hat Developer · 52 upvotes · 1 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#genai](https://daily.dev/tags/genai), [#diffusion-models](https://daily.dev/tags/diffusion-models) [View this post on daily.dev](https://daily.dev/posts/training-krea-2-what-matters-in-generative-model-training-sangwu-lee-krea-ai-ocrqtpdvc) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai","url":"https://daily.dev/posts/training-krea-2-what-matters-in-generative-model-training-sangwu-lee-krea-ai-ocrqtpdvc","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/training-krea-2-what-matters-in-generative-model-training-sangwu-lee-krea-ai-ocrqtpdvc"},"datePublished":"2026-08-18T14:20:02.278Z","dateModified":"2026-08-18T14:25:31.937Z","description":"A Krea.ai researcher describes the training pipeline behind Krea 2, the company's recently open-sourced image foundation model, focusing on stylistic diversity...","image":"https://i.ytimg.com/vi/-tviRdpmHvs/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/-tviRdpmHvs/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/training-krea-2-what-matters-in-generative-model-training-sangwu-lee-krea-ai-ocrqtpdvc","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,genai,diffusion-models","timeRequired":"PT21M","video":{"@type":"VideoObject","name":"Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai","description":"A Krea.ai researcher describes the training pipeline behind Krea 2, the company's recently open-sourced image foundation model, focusing on stylistic diversity...","thumbnailUrl":"https://i.ytimg.com/vi/-tviRdpmHvs/sddefault.jpg","uploadDate":"2026-08-18T14:20:02.278Z","duration":"PT21M","url":"https://api.daily.dev/r/ocRqTpdVc","embedUrl":"https://www.youtube.com/embed/-tviRdpmHvs"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/training-krea-2-what-matters-in-generative-model-training-sangwu-lee-krea-ai-ocrqtpdvc#faq","mainEntity":[{"@type":"Question","name":"How does Krea 2's approach to image generation differ from ChatGPT and Nano Banana Pro?","acceptedAnswer":{"@type":"Answer","text":"Krea 2 prioritizes stylistic diversity and fast generation over the slow, highly consistent but mode-collapsed outputs typical of production models like ChatGPT and Nano Banana Pro. Those models tend to render the most generic, centered version of a subject to guarantee reliability, taking up to a minute or two per image, whereas Krea 2 aims for quick iteration across varied visual styles for creative exploration. daily.dev surfaces training approaches like this for developers comparing image generation models."}},{"@type":"Question","name":"How do you filter AI-generated images out of a training dataset for an image model?","acceptedAnswer":{"@type":"Answer","text":"Krea's team trained a large vision-language model to classify whether an image looks AI-generated, then distilled that classifier down into a small, cheap SigLip-sized model so it could run efficiently across billions of images. They avoided training on synthetic data entirely because it makes the resulting model imitate the aesthetic of models like ChatGPT or Nano Banana Pro rather than learning from real data. Developers curating training data for generative models can track techniques like this on daily.dev."}},{"@type":"Question","name":"What training pipeline stages are used for modern production diffusion models?","acceptedAnswer":{"@type":"Answer","text":"Production-grade diffusion models follow a pipeline inspired by LLM training: low-to-high resolution pretraining (from 256px up to 1K), mid-training on curated domain data, supervised fine-tuning, preference optimization using pairwise comparisons, GRPO-inspired reinforcement learning against reward servers, and training a separate small LLM to expand short user prompts into long, detailed prompts before feeding the diffusion model. daily.dev helps engineers building generative pipelines keep up with evolving diffusion training stages."}}]} ```