KREA

Further reading Diffusion model (Wikipedia) · Text-to-image model (Wikipedia) · Krea Realtime 14B: Real-Time, Long-Form AI Video Generation (Krea)

1. i.

   One pass at a time

   Each denoising step is a full run of the network, and it has to wait for the one before. Good samples used to take hundreds or thousands of model evaluations. DistillationTraining a faster student model to reproduce what a slower teacher model does. chips away at this: a student model learns to match its teacher in half the steps, and doing that again and again gets you down to as few as 4 steps without losing much quality. Consistency models go further and map noise straight to an image, with extra steps on offer if you want some quality back.

2. ii.

   Words are a blunt tool

   Try describing a complicated layout, a pose or a shape in a text prompt and you'll see the problem. Getting close to the picture in your head usually means editing the prompt, looking at the result and editing again. If every round means waiting for a render, that loop gets old fast.

3. iii.

   The AI look

   You know it when you see it: overly blurry backgrounds, waxy skin, boring composition. Part of the blame may lie with the data. LAION Aesthetics, a scorer commonly used to pick training images, favours blurry backgrounds, soft textures and bright images, and that bias ends up baked into the model. Tuning on existing preference datasets has side effects too, including sliding back toward the same look.

4. iv.

   Video that can't stream

   If you denoise a whole clip at once, you can't show any of it until all of it is done, so real-time streaming is out. The alternative is autoregressiveGenerating output one piece at a time, with each new piece built on the ones already made. generation: finish a frame, show it, make the next. The catch is that the model now builds on frames it made itself instead of real footage. That mismatch is called exposure biasThe gap that arises when a model trained on real data must, in use, build on its own imperfect output., and small flaws can snowball.

Further reading Progressive Distillation for Fast Sampling of Diffusion Models (arXiv) · Consistency Models (arXiv) · Adding Conditional Control to Text-to-Image Diffusion Models (arXiv) · Releasing Open Weights for FLUX.1 Krea (Krea) · Krea Realtime 14B: Real-Time, Long-Form AI Video Generation (Krea)

Further reading Releasing Open Weights for FLUX.1 Krea (Krea) · What Is Krea 2? Krea's Foundation Image Model (Krea) · Krea Realtime 14B: Real-Time, Long-Form AI Video Generation (Krea) · Krea Realtime: live AI canvas for images and video (Krea)

1. Step 1: Start from a raw base

   For FLUX.1 Krea, KREA started from flux-dev-raw, a pre-trained 12B-parameter diffusion transformer from Black Forest Labs that is free of the AI aesthetic. Most of a model's look, the thinking goes, is learned in post-trainingFurther training after the broad first stage, used to steer a model toward the outputs its makers want.. So it fine-tuned on hand-picked images, then ran several rounds of preference optimisation on data collected with a clear, opinionated art direction.

2. Step 2: Teach video to stream

   Krea Realtime 14B is distilled from Wan 2.1 14B using Self-Forcing, a way of turning video diffusion models into autoregressive ones. It trains the model on its own earlier outputs, the way it will actually run, which goes straight at exposure bias. One stage of training cuts sampling from about 30 steps to 4. It runs at 11 frames per second, 4 steps a frame, on a single NVIDIA B200 GPU.

3. Step 3: Make style a dial

   Krea 2 is KREA's first foundation image model built from scratch. It carries the style of a reference image into the output, lets you choose how strongly, and can blend styles together. In the Realtime canvas, rough strokes, simple shapes and blocks of colour are enough to steer what comes out.

Further reading Releasing Open Weights for FLUX.1 Krea (Krea) · Krea Realtime 14B: Real-Time, Long-Form AI Video Generation (Krea) · What Is Krea 2? Krea's Foundation Image Model (Krea) · Krea Realtime: live AI canvas for images and video (Krea)

• How long can a streamed video hold together?

  An autoregressive model feeds its own mistakes back in as context, and this gets worse the longer the context grows. The Self Forcing authors report quality that matches slower non-causal models, but a model that's fine on short clips can still fall apart on the endless generation a live tool needs.

• How do you measure a good image?

  Common benchmarks mostly check whether the model did what the prompt said: spatial relationships, object counts and the like. Older aesthetic scorers aren't good enough for today's models, and preference may be too personal to squeeze into one number anyway. How one model, or one metric, serves many tastes is still open.

Further reading Krea Realtime 14B: Real-Time, Long-Form AI Video Generation (Krea) · Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (arXiv) · Releasing Open Weights for FLUX.1 Krea (Krea)

diffusion models

Generative models that make an image by starting from random noise and removing it step by step.

latent space

A compressed numerical version of an image that a model works in instead of raw pixels.

text embedding

A list of numbers that a language model produces to represent the meaning of a prompt.

Distillation

Training a faster student model to reproduce what a slower teacher model does.

autoregressive

Generating output one piece at a time, with each new piece built on the ones already made.

exposure bias

The gap that arises when a model trained on real data must, in use, build on its own imperfect output.

post-training

Further training after the broad first stage, used to steer a model toward the outputs its makers want.

1. 1Diffusion model · Wikipedia
2. 2Text-to-image model · Wikipedia
3. 3Progressive Distillation for Fast Sampling of Diffusion Models · arXiv
4. 4Consistency Models · arXiv
5. 5Adding Conditional Control to Text-to-Image Diffusion Models · arXiv
6. 6Releasing Open Weights for FLUX.1 Krea · Krea
7. 7Krea Realtime 14B: Real-Time, Long-Form AI Video Generation · Krea
8. 8Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion · arXiv
9. 9What Is Krea 2? Krea's Foundation Image Model · Krea
10. 10Krea Realtime: live AI canvas for images and video · Krea