Replicate black-forest-labs/flux-schnell

Flux Schnell is the fast, cheap, permissively-licensed tier of Black Forest Labs' Flux family, and this node runs it through Replicate's cloud instead of on your GPU. If you want a Flux image in a couple of seconds and don't care to load 12GB of weights locally, this is the low-effort path: type a prompt, get an image back. As a Replicate node it needs a Replicate account and a REPLICATE_API_TOKEN, and you pay per generation - but there's no download, no VRAM requirement, and it runs on any machine.

Schnell ("fast" in German) is the step-distilled Flux tier. It targets 1–4 steps, so it's the quickest and cheapest of the three Flux tiers, and it's Apache 2.0 - the fully open one. The trade is fidelity: it's a notch below Dev, which is itself a notch below Pro. For drafts, batches, and "just give me something now," it's the right pick.

How it works

Because Schnell is heavily distilled, there's no CFG and no negative prompt - that machinery was trained out. You don't set steps or a guidance scale here either; the model is tuned to do its thing in a handful of steps. What you get instead is a small, clean set of controls: prompt in, image out, fast.

The inputs and outputs that matter

• prompt (required) - Flux takes natural language well, so write a real sentence, not a keyword soup.
• aspect_ratio (1:1) - pick from ratios like 16:9, 3:2, 9:16. Note it's a ratio, not width/height fields.
• num_outputs (1, up to 4) - batch a few candidates in one call.
• go_fast (true) - the optimized fast path. Leave it on unless you're chasing a marginal quality bump.
• megapixels (1) - output size, 1 or 0.25. Drop to 0.25 for thumbnails or cheaper drafts.

Output is a single IMAGE. output_format (webp/jpg/png) and output_quality control the file, seed fixes the result, disable_safety_checker turns off the content filter, and force_rerun skips Replicate's cache when identical inputs would return the same image.

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.

No Flux weights land on your disk - the model lives on Replicate. The only dependency is the replicate pip package. Set your token before launching ComfyUI: export REPLICATE_API_TOKEN="r8_..." (replicate.com/account/api-tokens).

Common issues & troubleshooting

Your negative prompt isn't doing anything - because there's no field for one. Schnell is distilled and has no CFG, so negative prompting simply doesn't exist on this tier. Steer with the positive prompt instead. If you genuinely need negatives, you're looking at a different model.

Faces look generic and samey. That's the well-documented "Fluxmaxxing" effect, and it's worse on the distilled tiers (Schnell and Dev) than on Pro. Skip beauty words like "beautiful" and "gorgeous," add specificity about the person, and vary the pose and setting. If you need real facial variety and quality, step up to flux-dev or flux-pro.

Output blurry or low on detail. Schnell is the quality floor of the Flux line by design. It's built for speed and volume; for a keeper, generate candidates here and re-run the winner through flux-dev or flux-pro.

A prompt got refused or an image came back censored. The safety checker runs by default. disable_safety_checker exists, but note Flux models ship censored at the weights level too, so it won't unlock everything.