Fal API Flux Dev with LoRA

This is the plain Fal API Flux Dev node with five extra sockets bolted on the side, and those five sockets are the whole reason to reach for it. Dev happens to be the one Flux tier that's genuinely forgiving to train LoRAs against - the guidance distillation that gives it its homogenized-face reputation also makes it tolerant of a scrappy, inconsistent single-subject dataset in a way the undistilled base model isn't. That's exactly the kind of LoRA fal.ai lets you point at a URL and load server-side, and this node is where it plugs in.

How the LoRA wiring works. You don't type a LoRA path into this node directly. Instead, you build one or more "Fal API Flux LoRA Config" nodes - each one takes a lora_url and a scale - and wire their LORA_CONFIG output into lora_1 through lora_5. All five slots are optional and independent, so you can run zero, one, or stack up to five at once, each with its own strength. This is the pack's whole answer to community LoRA stacking, and it matches the README's own "Using LoRA" walkthrough: add a LoRA Config node, set the URL and scale, connect it in.

The rest of the inputs are identical to plain Flux Dev: prompt (multiline - write it as a sentence, Dev's T5 encoder is built for natural language, not tags), width/height (1024 default, step 8), num_inference_steps (28 default, 1–100), guidance_scale (3.5 default, 0.1–40 - worth dropping toward 1.8–2.0 for portraits, since 3.5 is the value most often blamed for Flux's same-face problem), num_images (1–4), and enable_safety_checker. seed is optional, 0 for random.

Output is a single IMAGE.

Installing it. No separate steps from the rest of the pack - install via ComfyUI Manager (search "ComfyUI-Fal-API-Flux") or clone the repo manually into custom_nodes (use the real yhayano-ponotech URL; the README's own manual-install example has a stray placeholder in it) and pip install -r requirements.txt. Then the part every node in this pack shares and every troubleshooting thread traces back to: a config.ini file in the pack's folder with a [falai] section and your fal.ai api_key. Restart ComfyUI after.

Where LoRA setups actually break. The README's own troubleshooting list flags this directly: make sure your LoRA URL is correct and actually compatible with FLUX.1 [dev]. In practice that means lora_url needs to be a direct, publicly reachable link to a .safetensors file - a Hugging Face resolve/main/... link works, a bare CivitAI model page URL usually doesn't, since fal's backend has to be able to download the file itself with no browser session behind it. If a LoRA silently does nothing, check that the URL actually resolves to a raw file and not an HTML page first. On scale: it runs 0.1–2, but pushing much past the LoRA's trained-in strength of roughly 1 starts to look overcooked the same way it does in local ComfyUI LoRA workflows - treat values above 1.2–1.5 as "I know what I'm doing" territory, not a default to reach for.