Replicate ai-forever/kandinsky-2.2

Kandinsky 2.2 is a text-to-image model from the AI-Forever / Sber group - a Stable-Diffusion-era model with its own distinct look and, architecturally, its own twist: it uses an image-prior stage before decoding. It never became a ComfyUI mainstay the way SD did, which makes it a nice fit for this pack: through ComfyUI-Replicate you can run Kandinsky on Replicate's cloud and pull its output into a graph without setting up a whole separate model stack for a checkpoint you might only want occasionally.

How it works

The thing that makes Kandinsky different from vanilla SD is the two-stage design. A "prior" model first maps your text into an image embedding (CLIP-style), and then a diffusion decoder turns that embedding into pixels. That's why this node has two step counts - one for the prior, one for the main diffusion. You don't have to think about the plumbing; the node sends your prompt and settings to Replicate and returns an IMAGE.

Inputs and outputs that matter

Output is an IMAGE.

• prompt / negative_prompt - your text.
• width / height - chosen from a fixed list of sizes (384 up to 1536-ish), default 512. Kandinsky 2.2's comfort zone is lower than SDXL's, so don't expect clean 1024² the way you would from SDXL.
• num_inference_steps (75) - the main decoder steps. High by default; you can bring it down to trade a little quality for speed and cost.
• num_inference_steps_prior (25) - steps for the prior stage. Leave it unless you're experimenting.
• num_outputs (1) - batch up to 4.

seed fixes the roll, output_format picks webp/jpeg/png, and force_rerun forces a fresh generation.

How to install it

ComfyUI Manager: search ComfyUI-Replicate, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/replicate/comfyui-replicate
cd comfyui-replicate
pip install -r requirements.txt

Restart, and set your token first:

export REPLICATE_API_TOKEN="r8_************"; python main.py

Token: replicate.com/account/api-tokens.

Common issues

Image output, cleanly supported - no pack output caveats.

Two model-specific things worth knowing. First, resolution: Kandinsky 2.2 is a 2023-era model and looks best at the sizes it was trained around. Push it to very large dimensions and coherence falls apart faster than a modern model would. Second, that default of 75 steps is generous and directly drives your cost per run - if you're just exploring, drop it substantially; you rarely need the full 75 to judge whether a prompt is working.

And the pack basics: a missing REPLICATE_API_TOKEN is the first-run error, and each generation bills on Replicate. Kandinsky has a recognizable aesthetic that some people specifically want - that's the reason to reach for this node over a more modern generator; if you just want the best general image quality, newer models will serve you better.