Generate creative image concepts and train custom styles

Krea 2 is an open text-to-image model family for generating images from plain prompts. The public demo exposes two versions: Turbo for fast few-step generation, and Raw for more controlled generation and style training.

The practical angle is workflow choice. If you want to test campaign concepts, art direction, product moods, or visual styles quickly, Turbo is the easier first stop. If you want to build a custom style workflow, the Raw checkpoint is the one Krea recommends for training LoRAs, which are small style adapters that can be applied back to Turbo for faster generation.

Start with the public Krea 2 Space and choose Turbo. Use one detailed prompt that describes the subject, setting, lighting, framing, and mood in normal language. For a useful first test, generate the same idea a few times and watch whether the model keeps the scene coherent while changing style.

If you want to go deeper, use the official GitHub repo or the Hugging Face model pages. The open workflow supports Diffusers, SGLang, and Krea's official inference code. Local use is GPU-heavy, so the browser demo is the better first evaluation path before downloading the full checkpoints.

This is a large image model. Local use can require serious GPU memory, especially at high resolution, so treat the hosted Space as the first test path and use local setup only when you need repeatable batches or style training.

• Explore campaign, poster, product, or editorial image directions from text prompts.
• Generate fast visual drafts before moving into a heavier design workflow.
• Test how one prompt behaves across different styles and compositions.
• Train a reusable style adapter on Raw, then apply it on Turbo for faster iteration.
• Compare the public demo against your local image-generation stack before committing setup time.

Try the demo

View model page