Kandinsky2.2 Unet Decoder

This is the workhorse. The Prior Loader and its encoder nodes handle the "what do we want to draw" half of Kandinsky 2.2; the Unet Decoder is the "draw it" half - a full UNet denoising pass over latents, conditioned on the image embeddings the prior produced. If a Kandinsky 2.2 graph has a single most important node, it's this one.

How it works

Kandinsky 2.2's decoder is a classic UNet diffusion pass, but with two things you won't see in a Stable Diffusion sampler node. First, it's conditioned on image_embeds passed as added conditioning rather than CLIP text embeddings - which is why everything upstream has to go through this pack's PRIOR_LATENT type. Second, this UNet predicts noise plus a variance term (learned-variance style), so the node runs CFG on the noise, re-attaches the variance, and only then steps the scheduler. That's why its default guidance_scale of 4 is half the prior's default of 8 - they're different stages tuning different things.

It also honors the strength slider the way img2img expects: lower strength starts the denoise loop at a later timestep, keeping more of the source image from the Image Latents node; strength 1 means full denoise from pure noise.

Inputs and output

• decoder - the DECODER output from the Decoder Loader.
• latents - from Latents (text-to-image) or Image Latents (img2img), optionally after a Hint Combiner.
• image_embeds, negative_image_embeds - the PRIOR_LATENT pair from the text/positive encoders or Image Encoder.
• strength (default 1) - how much of the starting image survives; the img2img dial.
• num_inference_steps (default 20), guidance_scale (default 4), seed.

Output: LATENT - denoised latents, which go into the MovQ Decoder for pixels. This node is not an image output node.

Installing it

Manual clone, per the README (not in ComfyUI Manager's list):

cd ComfyUI/custom_nodes
git clone https://github.com/vsevolod-oparin/comfyui-kandinsky22
cd comfyui-kandinsky22
python -s -m pip install -r requirements.txt

Models into ComfyUI/models/checkpoints:

git clone --depth 1 https://huggingface.co/kandinsky-community/kandinsky-2-2-prior
git clone --depth 1 https://huggingface.co/kandinsky-community/kandinsky-2-2-decoder
git clone --depth 1 https://huggingface.co/kandinsky-community/kandinsky-2-2-controlnet-depth

Gotchas

• Every port is required. Missing an embed port or wiring the wrong latent source is the most common way a Kandinsky graph just silently produces garbage. The simple workflow even uses two positive text encoders - one for image_embeds, one encoding a "low quality, bad quality" phrase into negative_image_embeds.
• Depth needs the right decoder. If you attach a hint via the Hint Combiner, the Decoder Loader must be set to kandinsky-2-2-controlnet-depth, or the hint is ignored. The README says so in bold; treat that as prophecy.
• Slow is normal. An 80-step decoder pass at 768×768 on top of a 20-step prior is not quick, and models shuffle between GPU and CPU between stages. It's a 2023 model doing 2023 speeds.
• Dependency pin. Old diffusers commit + accelerate==0.27.2 in requirements.txt can break other custom nodes sharing the environment.

Kandinsky 2.2 is long dead in the community, but if you're here, you already found the interesting part - this node is where its distinctive painterly look actually happens.