Krea 2 GGUF

Krea 2 is a 12B diffusion transformer for high-fidelity image generation and design. It uses layerwise and refiner text-fusion blocks and is not based on Flux. This repository ships GGUF quants of the Raw and Turbo checkpoints.

Hugging Face repositorykrea2-gguf on Hugging Face

Example output

Prompt: A seaside beach in the autumn from above. In the foreground there is a seagull flying. A text 'Krea2 GGUF'. Deep indigo and teal palette, minimalist surrealism, high-angle aerial perspective, atmospheric gradient, poignant isolation, liminal space

Available Quantizations

Each quant is published for both the Raw and Turbo checkpoints.

Quant	Size	Raw file	Turbo file
Q4_0	7.74 GB	krea2_raw_bf16-Q4_0.gguf	krea2_turbo_bf16-Q4_0.gguf
Q4_1	8.47 GB	krea2_raw_bf16-Q4_1.gguf	krea2_turbo_bf16-Q4_1.gguf
Q5_0	9.20 GB	krea2_raw_bf16-Q5_0.gguf	krea2_turbo_bf16-Q5_0.gguf
Q5_1	9.93 GB	krea2_raw_bf16-Q5_1.gguf	krea2_turbo_bf16-Q5_1.gguf
Q8_0	13.56 GB	krea2_raw_bf16-Q8_0.gguf	krea2_turbo_bf16-Q8_0.gguf
Q8_CR	12.84 GB	Not published	krea2_turbo_bf16-Q8_CR.gguf

Model Input and Output

Checkpoints

Checkpoint	Steps	CFG	Notes
Turbo	4 to 8	0.0	Distilled, no CFG. Fast.
Raw	20 to 30	3.0 to 7.0	Full CFG, more control, slower.

Outputs

Field	Value
Output	Image

ComfyUI Setup

Load the model with the GGUF loader node from comfyui-gguf-reboot. The loader node gives the best performance. The plain city96 node pack does not support the Krea 2 architecture.

Dependencies

Place these files in the ComfyUI model folders before running a workflow. You also need the text encoder and VAE from the upstream Krea 2 release.

Component	Folder
Diffusion transformer	models/diffusion_models/ or models/unet/
Text encoder	models/clip/ or text_encoders/
VAE	models/vae/

Krea 2 has no workflow files in this repository. Use the ComfyUI template from the upstream Krea 2 release and swap in the GGUF loader node.

Workflow downloads

Import these JSON files into ComfyUI.

JSON fileText-to-Image workflow