FalDiffusion

FalDiffusion is the shortest possible path from a prompt to a FLUX image: it sends your text to fal.ai's hosted flux/schnell and hands back the picture. No checkpoint, no CLIP text encoder, no VAE, no VRAM of your own. If you've ever wanted FLUX-quality output without a card that can actually run FLUX, this is the node - you're renting fal's hardware instead.

How it works

Under the hood it's a single call: fal_client.submit("fal-ai/flux/schnell", ...) with your prompt, image_size locked to square_hd, and num_inference_steps set to your steps. When the job finishes it downloads the result and returns it as an IMAGE. That's the entire node.

Two inputs, both worth understanding:

• prompt - a single-line STRING. Notice it's not multiline like most prompt fields, so you're composing in one line.
• steps - 1 to 8, default 2.

FLUX.1 Schnell is Black Forest Labs' step-distilled model - Apache-2.0, the fast sibling of Dev. Distillation means it wants very few steps; the community rule of thumb is that quality approaches Dev around 4 steps, and 2 is the draft-speed sweet spot. It's also guidance-distilled, which is a polite way of saying there is no negative prompt on Schnell. This node doesn't even have a field for one, so don't go looking.

The tradeoffs, stated plainly

This is a cloud API node, not a local one. It costs money per run, your prompt goes to fal's servers, and the output is locked to square_hd - no width/height control, no seed input, no batch count. If you've got a 12GB+ card, running Schnell or Dev locally gives all of that back and more. Where this shines: zero-install prototyping, offloading from a weak machine, or a quick sanity check before committing to a big local run.

Install and the key gotcha

In ComfyUI Manager, search comfyui-tensorops. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/un-seen/comfyui-tensorops

Restart; fal-client installs with the rest of the requirements. No models to download. You'll need a fal.ai API key - and here's the specific trap for this pack: its config.py unconditionally sets os.environ["FAL_KEY"] = "" every time the pack loads, which wipes any key you exported in your shell before launching ComfyUI. The working move is to paste your FAL_KEY straight into config.py. Get that wrong and every run dies with a missing-token error that looks like fal's fault. It isn't.