Krea 2

Highlights

• RAW = full-step base model, good diversity and malleability, suited to LoRA training / fine-tuning
• Turbo = 8-step distilled version, the official recommendation for most people's everyday use
• LoRAs trained on RAW work with Turbo; this is key to how fast its ecosystem has grown
• Supports 1K–2K output (set megapixels to 2.0 in ResolutionSelector for 2K)
• Ships with 9 official style LoRAs (ink wash, halftone dot, neon, rain window, retro anime, etc.) at weight 1.0

Scope

• Good for: photographic, cinematic, strongly stylized images; also good as a LoRA training base
• Good for: workflows that want "one base model plus a large library of style LoRAs"
• Not for: Chinese text within images (not its strength)

Tips

• Train LoRAs with RAW, generate with Turbo; this is the official two-stage usage
• To change style, prefer attaching a LoRA rather than piling on prompt adjectives
• The style reference workflow requires separately downloading the int8 weights and the krea2_style_reference LoRA; don't attach them to ordinary Turbo weights

Cautions

• The community license has restrictions on commercial use; read Krea's license page before commercial use

Parameters

RAW

52 steps (full-step sampling)

Turbo

8 steps

Resolution

From 1K up to 2K

Text encoder

Qwen3VL-4B (the fp8_scaled version saves VRAM)

VAE

Qwen-Image VAE

Style reference

There is also a dedicated style reference workflow with int8 convrot weights and the krea2_style_reference LoRA

Deployment

16GB verdict

Very easy to run: fp8 or int8 Turbo weights are about 12–13GB, and with a 4B encoder it's in the "comfortable on 16GB" tier.

VRAM notes

You already have both krea2_turbo_int8_convrot and fp8_scaled on your machine; use them directly. For 2K output, set batch to 1.

Suggested starting point

Try Turbo 8 steps and ResolutionSelector at 1MP first; move to 2MP once stable.

Comparison

It's a direct competitor to Z-Image Turbo in the 8-step generation race: ZIT is 6B, Apache-2.0, 650k downloads; Krea 2 has larger weights (about 12GB for fp8) and a stricter license, but Civitai already has 162 fine-tunes (ZIT has 100), so community remixing is more active. For LoRA training choose Krea 2's RAW; for generation choose Turbo.