Global background decoration
Krea 2Krea 2
...

Krea 2 RAW vs Turbo: Which Checkpoint Should You Use?

Krea2 Editorial
Krea 2 RAW vs Turbo: Which Checkpoint Should You Use?

Choose Krea 2 RAW when the work is to adapt the model. Start with Krea 2 Turbo when the work is to generate images using an existing checkpoint. If you are training a style adapter and then using it in production, the two checkpoints can play different roles in the same project.

This comparison covers the downloadable checkpoints described in Krea's open-weights release. It is based on developer documentation, not a local speed or image-quality test.

Decision Krea 2 RAW Krea 2 Turbo
What it is Undistilled base checkpoint Distilled checkpoint designed for eight-step generation
Where to start Research, fine-tuning, or training a LoRA adapter Text-to-image inference or evaluating a compatible trained adapter
What to budget for Training time, dataset preparation, and evaluation Runtime setup, generation, and review of rejected images
Main decision risk Paying for training before defining a stable visual target Assuming faster generation will fix an inconsistent style brief

Start with RAW when the style needs to live in an adapter

Imagine an editorial team creating a series of ceramic product concepts. If the team still disagrees on the palette, materials, or linework, more training will not settle the art direction. First approve a small set of examples and identify traits that should remain consistent. The creative exploration guide is a starting point if that visual brief is still open.

A clay master form beside ceramic sample tiles in a workshop

Generated concept illustration for exploring a range of treatments. No checkpoint output is shown.

Once the style is defined, an adapter can make sense if it must work across many different subjects. Krea recommends training LoRAs on RAW and applying them on Turbo. Confirm support in your runtime before committing to that handoff, and keep a few prompts outside the training examples for evaluation. If the adapter repeatedly recreates a training subject when asked for another object, treat that as a failure even when the surface treatment looks right.

Start with Turbo when you need images now

For a one-off design review or an initial batch, first try generation without training. Turbo is designed for inference with fewer steps, but its step count is not a promise about seconds per image. Hardware, image size, runtime setup and queueing can all change the time you experience.

Rows of cobalt ceramic cups on a long worktable

Generated metaphor for repeated inference, not a throughput measurement.

If you only need to explore prompts in a browser, the Krea2 image workspace offers a way to prototype with the models and credit settings displayed there. This comparison does not assume that the workspace exposes the downloadable RAW or Turbo checkpoints. For hosted Krea API rates, check the developer's API pricing separately. A hosted per-image rate does not tell you what a local training run or a self-hosted batch will cost.

Use a small acceptance set before scaling

Pick three prompts: one close to approved work, one with a different subject, and one with a difficult composition. Keep the output size and review criteria consistent. If you are testing an adapter, compare Turbo with it enabled and disabled. Record the checkpoint and adapter versions, runtime settings, and which images pass. When comparing two checkpoints, use each one's documented settings; forcing both to use the same number of steps may answer the wrong question.

Hands comparing ceramic design proofs next to a clay cup

Generated illustration of a review session, not a side-by-side model test.

Measure cost per accepted image, not just cost per generated image. For example, if a hypothetical run costs $6 and yields 12 acceptable images, its generation cost is $0.50 per accepted image before review time. Include training and setup separately if they apply. Review the Krea 2 Community License for the intended deployment and distribution before relying on the downloadable weights; access to files does not settle the licensing question.

If unadapted Turbo meets the brief, there is no reason to train simply because RAW is available. If a stable style keeps breaking across new subjects, test an adapter trained on RAW, validate it on Turbo, and compare the number of usable images before expanding the run.