RunPod Flux Schnell (Fast)

Schnell is Flux's step-distilled little brother, and this node is it as a RunPod public endpoint: you get a fast, cheap, Apache-2.0 Flux without a local GPU or a local model file. The whole pitch is in the default - 4 inference steps. Where its sibling Flux Dev runs 28 steps and waits a minute or two, Schnell finishes in four and is built to be a draft stage, not a final render.

That's the honest frame for it. Schnell is more aggressively distilled than Dev, which the community notes is exactly why Black Forest Labs felt safe releasing it under Apache 2.0 - it's lower quality and nobody's going to finetune it into something better. You use it to iterate on composition and idea space, then take the winning prompt to a bigger model. This node is the version that costs pennies instead of a GPU.

How it works

Identical plumbing to the rest of this pack: the node POSTs your prompt to black-forest-labs-flux-1-schnell on RunPod's API, polls the async job until it completes, and decodes the result into a ComfyUI IMAGE tensor. You lose nothing workflow-wise - it behaves like a local sampler that happens to live in someone else's data center.

The inputs that matter

• api_key - your RunPod key; paste or share via a Primitive.
• prompt - the generation text. This is where you do your iterating.
• num_inference_steps - default 4, max 8. You can push to 8 for a bit more coherence, but you're fighting the point of the node.
• guidance - default 7.5 here too, and like its Dev sibling that's high. Schnell is fully distilled - traditional guidance barely applies - so don't expect big swings from this knob.
• width / height - 256 to 1536 in 64 steps, default 1024.
• negative_prompt - sent, largely ignored; fully distilled models don't do CFG. Don't build your workflow around it.
• seed - -1 for random, set it to lock a composition.
• image_format - default jpeg here (vs png on Dev), which tells you how the endpoint thinks of this node: drafts, not gallery pieces.

What comes out

One image output (IMAGE). Preview it, or skip straight to Save Image - Schnell is for volume, so wire it to something that doesn't make you think too hard.

Installing it

Same pack, same install as every node here: search ComfyUI_RunpodNodes in ComfyUI Manager, or

cd ComfyUI/custom_nodes
git clone https://github.com/tcarwash/ComfyUI_RunpodNodes

then restart. No extra Python dependencies - the pack only needs what ComfyUI already has. Grab a RunPod API key at runpod.io.

Gotchas

• It's a draft tool, not a final tool - at 4 steps you get Fast Flux, not finished art. If you're reaching for Schnell for hero images, you're using the wrong node.
• Still metered - it's fast and cheap, but it's not free. For serious prompt-soup iteration, a free hosted schnell endpoint or your own GPU still beats paying per call.
• "Job failed" - key or credits, as always with this pack. Verify both before anything else.
• Expect queue time anyway - four steps of compute don't skip the public-endpoint queue; a busy server still costs you 30+ seconds of waiting. The speed is in the model, not the logistics.

The honest use case: you're on a no-GPU machine, or your GPU is busy, and you want Flux-quality composition feedback in volume. Schnell via this node does exactly that. Just remember it's the sketchbook, and save the gallery for Dev or a local run.