Ana içeriğe geç

Kandinsky 2.2

Model Source Revision Weight license Parity (CPU, fp32)
kandinsky-2.2 kandinsky-community/kandinsky-2-2-decoder 9ae140d Apache-2.0 Verified

Catalog: kandinsky-2 · Models: kandinsky-2.2 · Task: t2i

Architecture. A CLIP text model with projection conditions a PriorTransformer, while a CLIP vision model with projection supplies the zero image embedding for the negative branch. The prior uses an UnCLIP scheduler with sample prediction and fixed_small_log variance. Its image embedding conditions a U-Net with ResnetDownsampleBlock2D, SimpleCrossAttn* blocks and scale_shift time modulation. A MoVQ VQ model with spatially conditioned normalization decodes the latents. The decoder uses a DDPM scheduler with linear betas and clipping.

Usage.

python -m imagehub.pretrained kandinsky-2.2 checkpoints/kandinsky-2.2
from imagehub import pipeline

with pipeline(
    "t2i", backend="kandinsky-2.2", model="checkpoints/kandinsky-2.2", device="mps"
) as pipe:
    result = pipe("a cat, oil painting", seed=42)
    result.images[0].save("cat.png")

Settings. The prior defaults to 25 steps and guidance 4.0; the decoder defaults to 100 steps, guidance 4.0 and 512×512 output. The supported options are steps, guidance, prior_steps, prior_guidance and negative_prompt. 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.

Out of scope. Img2img, inpainting, ControlNet, prior interpolation and fp16/bf16 parity are not supported.

License. The decoder and prior weights use the Apache-2.0 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.