Krea 2

Krea 2 by Krea AI: a picture model trained from scratch, with an architecture of its own. The app shows the family as Krea 2.

#What it is

Krea 2 is not a version of Flux. It comes in two kinds:

• RAW: the base, made for training LoRAs; it needs many steps (about 52).
• Turbo: distilled to 8 steps, for making pictures.

#Not FLUX.1 Krea [dev]

FLUX.1 Krea [dev] is another model with a similar name: a Flux.1 dev model trained again by Krea. It loads exactly like Flux.1 and belongs to the Flux.1 family. The app sorts files the same way: a name that holds both Flux and Krea ("Flux.1 Krea", "krea-flux") is Flux.1; a name that holds only "krea" or "krea2" is Krea 2.

#Files and nodes

• The model, for example krea2_turbo, in models/diffusion_models, with Load Diffusion Model.
• The text encoder qwen3vl_4b in models/text_encoders, with Load CLIP, type krea2.
• The VAE qwen_image_vae in models/vae, with Load VAE.

Start from ComfyUI's Krea 2 template for the exact files.

#Settings that work

The template of Krea 2 Turbo: Empty Latent Image 1024 × 1024, KSampler 8 steps, cfg 1, euler, simple, and ConditioningZeroOut in place of a negative prompt. RAW needs about 52 steps.

#What fits

Only files made for Krea 2. Flux LoRAs do not work with it. LoRAs trained on RAW also work on Turbo.

#How the app helps

Shown as Krea 2 on model cards. Hold the pointer over the family label to read the built-in note: what Krea 2 is, where its files go, and that it is not FLUX.1 Krea [dev], so Flux LoRAs do not work with it.

#See also

• Flux.1
• Model families
• Why families do not mix
• Separate loaders