Krea 2's workhorse components on both ends turn out to be Qwen, spec dive reveals

Even_Package_8573 · reddit · 2026-09-24

The author read Krea 2's spec instead of the sample grids flooding feeds and found that while the backbone is trained from scratch, the two parts doing the real work are both Qwen components: the text encoder is Qwen3-VL 4B, which takes twelve decoder layers per token and fuses them inside the transformer rather than just reading the last hidden state; the Qwen-Image VAE decodes the latent back to an image.
If you already know how Qwen-Image handles text and composition, you mostly understand Krea 2's behavior.
Clarification on Raw vs Turbo: Raw is not an unfiltered checkpoint — it's the undistilled mid-training base for fine-tuning and LoRA work (28 steps, CFG 4.5, 26GB for the diffusers install). Turbo is the generation model: 8 steps, guidance off. Complaints about heaviness come from using a training base for portraits.
Recommended workflow: train LoRAs on Raw with 50-60 image sets, generate with Turbo; quantized Turbo for casual image generation.

More from Models