Ana içeriğe geç

Krea 2

Model Source Revision Weight license Parity (CPU, fp32)
krea2-turbo krea/Krea-2-Turbo 98e0fe1 Krea 2 Community License Verified

Catalog: krea-2 · Models: krea2-turbo · Task: t2i

Architecture. A Qwen3-VL text model with 36 layers and width 2560 supplies the hidden states of 12 selected layers, stacked for a roughly 13B diffusion transformer with 28 blocks, 48 query heads and 12 key-value heads of width 128. The transformer uses 3-axis RoPE plus layerwise and refiner text blocks. A Qwen-Image VAE uses 16 latent channels, causal 3D convolutions and per-channel latent mean and standard deviation. The flow-match Euler scheduler uses exponential dynamic shifting. The native image size is 1024×1024.

Usage. The repository is gated. Accept its terms and provide HF_TOKEN when downloading:

HF_TOKEN=... python -m imagehub.pretrained krea2-turbo checkpoints/krea2-turbo
from imagehub import pipeline

with pipeline("t2i", backend="krea2-turbo", model="checkpoints/krea2-turbo", device="mps") as pipe:
    result = pipe("a cat, oil painting", seed=42)
    result.images[0].save("cat.png")

Settings. The model defaults to 8 steps, guidance 0.0 and 1024×1024; width and height must be multiples of 16. The other supported options are negative_prompt, seed, latents and max_sequence_length. Unrecognized settings raise an error.

Seed compatibility. On CPU, the same seed and settings give the same image as diffusers with generator=torch.Generator("cpu").manual_seed(seed) (validation). MPS and CUDA runs are smoke tests only; there is no claim of exact equality on these devices. The published bfloat16 weights are upcast for float32 parity.

Out of scope. Krea 2 Raw, the single-file turbo.safetensors in the official codebase layout and fp16/bf16 parity are not supported.

License. The weights use the Krea 2 Community License; see the license. imagehub does not distribute weights.

Reference and imagehub, side by side

Drag the slider. Left is the recorded reference, right is imagehub, from the same seed, on CPU in float32.