Flux Schnell (fal)

Schnell is the Flux you reach for when you're fishing, not framing. It's the distilled tier of FLUX.1 - trained to land roughly where Dev ends up after 28 steps in just 4 - so it's fast and cheap, at the cost of some fine detail. FluxSchnell_fal runs that model on fal.ai's hardware and returns a native IMAGE, so you get the speed win without owning a GPU big enough for a 12B model. Per-call it's the cheapest way into the Flux family on this pack.

The mental model the community settled on: distilled for seed farming and iteration, full model for the final render. Batch a dozen Schnell rolls to find a composition and seed you like, then re-render that idea on FluxDev_fal or a Pro node when you want the polish.

Why there's no guidance slider

Look at the inputs and you'll notice something missing: there's no guidance_scale. That's not an oversight. Schnell is guidance-distilled - the classifier-free guidance is baked into the weights during training, so there's no CFG knob to turn and no true negative prompt. Feeding a distilled model normal CFG would apply the correction twice and give you oversaturated, high-contrast junk. The 4-step count is trained-in too; cranking num_inference_steps way up doesn't buy quality, it just costs more and can add artifacts. Leave it near the default.

The inputs that matter

• prompt - natural-language sentences, same as the rest of Flux. Schnell's a little less faithful than Dev, so keep prompts direct.
• num_inference_steps - defaults to 4. That's the sweet spot the model was built for. Nudge it a hair if you want, but this isn't the dial that fixes a bad image.

The rest is routine: image_size presets plus width/height, num_images (up to 10 - handy since Schnell is cheap enough to batch), seed (-1 for random), and enable_safety_checker. One IMAGE output, straight into Preview or Save Image.

Installing it

Easiest path is ComfyUI Manager → search ComfyUI-fal-API → install → restart. Manual install:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
pip install -r ComfyUI-fal-API/requirements.txt

No models to download - the whole pack is API glue. Set your fal key once: copy config.ini.example to config.ini and paste the key from fal.ai/dashboard/keys, or export FAL_KEY=your_key. Restart, and the node lands under FAL.

Where people get burned

The usual one is judging Schnell by Dev's standards. It will not match Dev on skin, hands, or micro-detail - that's the distillation tax, and it's inherent, not a settings problem. If a Schnell image looks soft, that's the model, not a bug; move the winning seed to Dev.

Second: it's still a paid API call. It's cheaper than Dev, not free. The pack's persistent cache means identical prompts don't re-bill, so use it while you're iterating and toggle force_rerun only when you want a genuinely new roll.

And if a call errors, read the message - the pack surfaces fal's actual error (quota, content policy, a bad parameter) instead of silently returning nothing. Schnell being Apache-2.0 licensed upstream doesn't change that fal runs its own safety checks on the hosted endpoint.