Fast iteration changes how you explore visual ideas — Krea 2 Turbo compresses Krea 2's sampling into an 8-step distilled checkpoint so you can produce style-aware images in seconds rather than long denoising runs. The tradeoff is deliberate: Turbo favors speed and consistent style over the raw checkpoint's full malleability, making it ideal for concepting and prompt testing.
Key Capabilities
- 8-step distilled sampling optimized for quick generations (reported ~4 seconds per generation in hosted settings) — so you can iterate many prompts and directions rapidly without high compute or long waits.
- Native support for style references and moodboards (up to 4 images) and seed control — so designers can guide consistent visual language across batches.
- LoRA/Raw workflow compatibility: train LoRAs on the undistilled Krea 2 Raw checkpoint and apply them to Turbo for inference — so fine-tuning remains flexible while inference stays fast.
- Practical resolution and sampling recommendations: works well at 1K–2K outputs; recommended Turbo settings include few-step sampling (8 steps), CFG can be disabled for Turbo runs, and a timestep-shift mu (e.g., ~1.15) is suggested for best visuals — so you get predictable, high-quality outputs when using the suggested sampler settings.
Who it's for and trade-offs
Great fit if you need rapid visual exploration, style-driven concept batches, or low-cost iteration for illustrations, graphic design, and creative exploration. Use Turbo to validate directions quickly, then move promising results to larger Krea variants or a Raw-based finetune if you need more diversity or ultra-high fidelity. Look elsewhere if you need a fully malleable base for training (use Krea 2 Raw), maximum sampling diversity, or the absolute highest-resolution/polish (use Krea 2 Medium/Large). Also note this is an open-weights community-licensed release: deployers are expected to implement content filtering and safety measures per the license, and distilled checkpoints may reduce some generative diversity compared with the Raw checkpoint.
Where it fits
Turbo sits at the fast-iteration end of the Krea 2 family: train and customize on Raw, then run high-throughput inference on Turbo. It fills the gap between prototype-only models and heavier production-grade large checkpoints by prioritizing speed and consistent stylistic control.
