Replicate fofr/sd3-with-chaos

This is Stable Diffusion 3 with a chaos knob bolted on - fofr's take on "give me SD3, but make the outputs less samey." Instead of the usual careful, deterministic run, it injects controlled randomness so a single prompt sprays out wildly different compositions. It's built for exploration: mash generate, see what weird thing comes back, keep the good seeds. If you've ever felt a modern model gives you the same three ideas over and over, that's the itch this scratches.

The catch you should know up front: this isn't running on your machine. It's a Replicate node. Your prompt goes to Replicate's cloud, an SD3 model runs there, and an image comes back. You need a Replicate account, a REPLICATE_API_TOKEN, and you pay per generation. No weights download, no VRAM used.

How it works

fofr is one of the more prolific model-builders on Replicate, and "chaos" models are a signature move - wrap a base model, add parameters that push it off its comfortable path. Here the base is SD3, Stability's MMDiT architecture. The architecture itself was a genuine contribution (the whole field followed its rectified-flow direction), even though the open 2B release famously flopped. The hosted version this node calls doesn't carry that baggage - you're getting a working SD3 endpoint with variety cranked up.

The inputs and outputs that matter

The whole point lives in two controls:

• chaos (default 5, range 0–10) - how much randomness to inject. Low is closer to plain SD3; high throws structure to the wind. This is the dial you'll actually play with.
• weird (default false) - a boolean that pushes things further into strange territory. Flip it on when normal chaos isn't strange enough.

Beyond those: prompt (write plain descriptive language), aspect_ratio (1:1 by default, plus ratios like 16:9, 2:3, 9:16), guidance_scale (4.5), and number_of_images (up to 10 at once - handy, because exploration wants volume). output_format and output_quality set the file; force_rerun forces a fresh generation instead of a cached one.

Output is a single IMAGE you wire straight into a Preview or Save node.

How to install it

One install covers 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.

Nothing lands on your GPU - the model runs on Replicate's servers, so the only thing that installs is the replicate Python client. Before launching ComfyUI, set your token: export REPLICATE_API_TOKEN="r8_..." (grab one at replicate.com/account/api-tokens). No token, no runs.

Common issues & troubleshooting

"Authentication failed" or the node errors on run. Almost always the token. REPLICATE_API_TOKEN has to be set in the environment before ComfyUI starts - set it, restart, try again.

Same image every time. Replicate caches identical inputs, so an unchanged prompt returns the earlier result. Bump number_of_images, change the prompt, or flip force_rerun on to force a genuinely new run.

Chaos is too much (or not enough). That's the chaos slider doing its job. Dial it down toward 2–3 for coherent-but-varied, up toward 8–10 for genuinely unhinged. Turn weird off first if outputs are unusable.

It costs money and I forgot. Every run is a billed Replicate prediction, and number_of_images: 10 is ten of them. Great for exploration, easy to rack up - keep an eye on your Replicate usage.