# Training Krea 2: What Matters in Generative Model Training Sangwu Lee, Krea.ai | AI Engineer World's Fair 2026 | 21:46 Source: https://www.youtube.com/watch?v=-tviRdpmHvs Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/training-krea-2-what-matters-in-generative-model-training Published: 2026-08-18 Tags: data-pipelines, distillation, image-generation, reinforcement-learning ## TL;DR - Krea 2 prioritizes stylistic range and fast generation so creative teams can explore before they know exactly what they want. - After the architecture is fixed, Sangwu Lee says data curation has the largest effect on image model quality. - Krea's pipeline filters, deduplicates, captions, rebalances, and trains on data at billion-image scale without using AI-generated images. ## Summary Sangwu Lee explains how Krea trained Krea 2 and why the team focused on stylistic diversity rather than the highly consistent outputs of slower production models. He gives a practical account of data work: removing duplicates, detecting artifacts, improving captions, preserving unusual aesthetics, checking coverage of world knowledge, and filtering AI-generated images. A painting photographed against a white wall becomes a useful example of how captioning errors can teach the model an unwanted visual rule. He then describes a training pipeline that moves from low-resolution pretraining to higher resolutions, mid-training, supervised fine-tuning, preference optimization, reinforcement learning, and prompt expansion. Lee says fast iteration depends on infrastructure, simple and scalable methods, efficient training, and reusing ideas from language-model research. He ends with possible directions including multi-expert distillation, removing separate components from the stack, and using richer visual representations such as bounding boxes and scene graphs. ## Key ideas ### Krea 2 trades some production consistency for faster exploration and wider styles [00:52](https://www.youtube.com/watch?v=-tviRdpmHvs&t=52s) Sangwu Lee says large production models often generate slowly but produce reliable, acceptable images with few visible flaws. Their consistency can come with mode collapse. For example, a model asked for a person may produce a centered, average-looking person, while repeated prompts for a burning skull produce outputs that are competent but barely different. Krea 2 focuses on faster generation and more stylistic diversity. Lee says this fits creative studios that have not decided what they want yet and need to explore several visual directions. A highly controlled model is useful for a specific poster or birthday card, but it is less useful during open-ended visual development. ### Latent diffusion makes image training cheaper by avoiding direct pixel modeling [03:23](https://www.youtube.com/watch?v=-tviRdpmHvs&t=203s) Lee gives a short explanation of diffusion training. The model sees an image with noise added and learns how to remove that noise. Generation starts from random noise and repeatedly removes it until an image appears. Krea 2, like many competitive diffusion models, works in a compressed latent space produced by an autoencoder instead of modeling every raw pixel. Lee connects this choice to the cost of transformer computation. Non-sparse transformers have roughly quadratic time complexity, so modeling every pixel directly becomes expensive. Spatial compression lets the model learn in a smaller representation and then decode the result back into an image. ### Once the architecture is fixed, data curation determines much of the model's behavior [05:07](https://www.youtube.com/watch?v=-tviRdpmHvs&t=307s) Lee repeats that data becomes the main source of improvement after the architecture is locked. Krea wanted stylistic diversity, so it avoided filtering data only by conventional image-quality scores. A low-resolution CRT video may look poor by those measures but still contain an aesthetic that users want. The team also treats duplicates, overrepresented concepts, caption failures, and images that are too crowded for a given resolution as data problems. Lee says curation must avoid removing unusual styles by accident. At low resolution, for example, an image containing thousands of characters may be too difficult for the model to learn, even if the image itself is otherwise valid. ### Krea rejects AI-generated images because their style is easy to inherit and hard to remove [07:20](https://www.youtube.com/watch?v=-tviRdpmHvs&t=440s) Krea tries to remove AI-generated images from its training data. Lee acknowledges that synthetic data can provide a shortcut to a good-looking model, but says its aesthetic is sticky. A model trained heavily on outputs from systems such as ChatGPT image generation or Google's image products can inherit their visual character. He says he can often recognize when a model has been heavily distilled from another production model. That defeats part of the research goal because the new model becomes an imitation of an existing system. Lee also says the issue affects his motivation as a researcher when the work amounts mainly to distillation. ### Caption failures can teach the model visual rules that users did not ask for [08:16](https://www.youtube.com/watch?v=-tviRdpmHvs&t=496s) Krea's captioning pipeline first runs OCR to extract visible text, adds metadata such as the identity of a famous person when known, and then uses a vision-language model to write a detailed caption. The caption can later be rewritten into JSON prompts or another training format. Lee's painting example shows why this matters. Captioners repeatedly described the subject of a painting while leaving out that it was framed and photographed on a white wall. When such examples enter training, a request for a painting can cause the model to place it on a white wall. Krea therefore designs filters for images whose important visual features are not captured consistently, or lowers their sampling rate. ### Billion-image filtering requires cheap stages before expensive semantic checks [10:14](https://www.youtube.com/watch?v=-tviRdpmHvs&t=614s) Krea may train on two to ten billion images, so it begins deduplication with inexpensive hash-based methods such as perceptual hashes or MD5 hashes. After reducing the set, it uses embedding methods such as SSCD and image-text embeddings to find semantic or near duplicates. For other filters, a large vision-language model first makes judgments such as whether an image appears AI-generated. Krea then distills that decision into a smaller classifier, making it cheap enough to run across a billion-image corpus. Lee compares this with work on large text datasets, where a large model creates taxonomies or quality judgments that are distilled into a smaller model for broad filtering. ### Sparse autoencoders provide unsupervised tags for artifacts and unwanted patterns [11:56](https://www.youtube.com/watch?v=-tviRdpmHvs&t=716s) Lee describes using sparse autoencoders trained on vision-model representations as an unsupervised tagging system. An image can activate sparse features associated with concepts such as a horse, black and white, blur, or a blurry image. Those features can then be used to find data to remove or oversample. Lee gives watermarks, signatures, and border artifacts as examples of patterns that may be caught this way. This approach avoids requiring a manually labeled dataset for every filter. It turns the features learned by a vision model into searchable signals for dataset curation. ### The training stack follows language-model practice from pretraining through preference and reinforcement learning [15:10](https://www.youtube.com/watch?v=-tviRdpmHvs&t=910s) Krea trains from 256-pixel images up to 1K resolution, starting with semantic learning at low resolution and adding structure and detail later. After pretraining, mid-training and supervised fine-tuning shape the model toward uses such as illustration, graphic design, photography, and cinematics. Preference optimization uses pairs of images to make the model more opinionated about the desired behavior. Reinforcement learning sends generated images to reward servers, which provide feedback on areas such as text rendering and anatomy. Krea also trains a prompt expander that turns a short user prompt into a longer description closer to the model's training distribution. Lee says the team is exploring multi-expert on-policy distillation for capabilities such as photography and text rendering. ## Notable quotes - "The easiest and most reliable way to render a person is render the most boring average person that exists and then put it in a center frame." (02:13) - "Really, data is quite everything that goes into the model." (05:07) - "Synthetic data is so sticky to the model that once you start training on AI image data, sure your model is good, but you kind of lose the point." (07:40) - "If you have good data, that's going to be valuable no matter what the hot new training paradigm is." (18:43) - "I really like to simplify the stack so that we can get rid of VAE and then text encoders and then just train a single clean transformer." (19:58) ## Tools & references mentioned - Krea 2 - ChatGPT - Nano Banana Pro - SSCD - CLIP - Essential Web Data - Wikipedia - Barack Obama the horse - DALL-E 2 - Ideogram - FAIR - Yann LeCun's lab ## Who should watch - You are training an image model and need concrete methods for cleaning and balancing data at very large scale. - Your model looks polished but too similar to existing production systems, and you want to understand how captioning and synthetic data affect that result. - You are building a diffusion training stack and want to compare its stages with language-model pretraining, fine-tuning, preference optimization, and reinforcement learning.