Replicate ai-forever/kandinsky-2

Kandinsky 2 is a text-to-image model with a distinctly dreamy, painterly aesthetic and an unusual architecture under the hood. It came out of the AI-Forever / Sber group and had its moment before SDXL and Flux took over. Today it's mostly a "different look" tool - you reach for it when you specifically want its style, not because it beats a modern model on sharpness or prompt adherence. This node runs it through Replicate's cloud, so there's nothing to download and no VRAM needed; you just need a Replicate account, a REPLICATE_API_TOKEN, and to accept that each run costs Replicate credits.

How it works

The interesting thing about Kandinsky is that it's a two-stage model. A "prior" first maps your text prompt into an image embedding (a CLIP-style latent), then a diffusion decoder turns that embedding into pixels. That's why you'll see two sets of controls: the prior_* ones tune the text-to-embedding stage, and the sampler ones tune the pixels stage. Most models fold this into one step; Kandinsky exposes both, which is unusual and occasionally useful.

The inputs and outputs that matter

• prompt - what to generate. Kandinsky likes fairly direct descriptions.
• num_inference_steps (50) and guidance_scale (4) - the decoder's step count and prompt adherence. Note the default guidance is lower than SDXL's; that's tuned for this model, don't crank it blindly.
• width / height (512) - pick from the offered sizes (256 up to 1024). It's a 512-native-era model, so very large sizes aren't its strength.
• prior_cf_scale (4) and prior_steps ("5") - the prior stage's guidance and steps. Leave these near defaults unless you're deliberately experimenting.
• scheduler - p_sampler (default), ddim_sampler, or plms_sampler.

Output is a single IMAGE (batch up to 4 via batch_size). output_format/output_quality set the file, seed fixes the result, and force_rerun bypasses Replicate's cache.

How to install it

One install brings every Replicate node.

• ComfyUI Manager: search ComfyUI-Replicate, install, restart.
• Manual:

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

  then restart.

No weights download - just the replicate pip package. Set your token first: export REPLICATE_API_TOKEN="r8_..." (replicate.com/account/api-tokens).

Common issues & troubleshooting

It's not as sharp or prompt-accurate as SDXL/Flux. Right - it's an older model, and that's the honest baseline. Use Kandinsky when you want its particular painterly, slightly surreal look; if you want maximum fidelity and prompt adherence, this isn't the model, and no setting will close that gap. It's a stylistic choice, not a quality upgrade.

Big images look off. Kandinsky is happiest around its 512-era resolutions. Generate at a supported size near 512–768 and upscale afterward rather than asking for 1024 directly.

Changing guidance_scale made it worse. The default of 4 is deliberately lower than the CFG 7+ you'd use on SDXL. Kandinsky over-bakes if you push guidance too high - small adjustments only.

Same image on a re-run. Replicate returned the cached result for identical inputs. Change the seed or set force_rerun.