Flux LoRA (fal)

This is Flux Dev plus LoRAs, run on someone else's GPU. FluxLora_fal is the text-to-image node for when you want a specific character, style, or concept baked in via a LoRA but don't want to download the 12B base or the adapters. You give it a prompt and up to two LoRA URLs; fal loads the base, applies the adapters, renders, and hands back a native IMAGE. No local weights, no VRAM ceiling.

A LoRA is a small low-rank adapter - usually tens to a couple hundred MB - that patches a checkpoint's behavior without retraining it. It's the unit of currency of every model ecosystem: one file adds a face or a look. This node is how you spend that currency against Flux without hosting anything yourself.

The one rule that bites everyone

LoRAs are architecture-bound. A Flux LoRA works here; an SDXL, Pony, or Illustrious LoRA does not, no matter how similar the file looks. If your lora_path points at the wrong architecture, you either get an error or an image where the LoRA plainly didn't take. Point these slots at Flux LoRAs specifically - a Hugging Face repo, a CivitAI download, or a fal-hosted .safetensors URL.

The inputs that matter

• prompt - plain sentences, like all of Flux. If your LoRA has a trigger word, the community consensus is to place it inside a sentence rather than prepend it as a bare tag; the T5 encoder handles it better that way.
• lora_path_1 / lora_path_2 - the URL of each LoRA. Two slots, so you can stack a character LoRA with a style LoRA.
• lora_scale_1 / lora_scale_2 - strength, default 1.0, range 0–2. The old "0.5–0.8" rule of thumb is SDXL advice; on Flux many LoRAs want to sit near 1.0. Turn it down if the LoRA is overpowering the scene.

Standard Flux controls round it out: image_size/width/height, num_inference_steps (28), guidance_scale (3), num_images, seed, enable_safety_checker. One IMAGE output.

Installing it

ComfyUI Manager → search ComfyUI-fal-API → install → restart, or:

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

No downloads - even the LoRAs live at their URLs, not on disk. Set your fal key (copy config.ini.example to config.ini, or export FAL_KEY=your_key; keys at fal.ai/dashboard/keys), restart, and look under FAL.

Where people get burned

Besides the architecture mismatch above, the common snag is a LoRA URL fal can't reach - a gated Hugging Face repo, or a link that isn't a direct download. It has to resolve to the actual .safetensors. If a LoRA "does nothing," check the scale isn't at 0, then check the URL actually points at a Flux adapter.

Trained your own LoRA on fal with FluxLoraTrainer_fal? Its output is a lora_file_url - feed that straight into lora_path_1 here and you've closed the loop entirely on fal's hardware. And as always, each render is a paid API call; the pack's result cache keeps identical prompts from re-billing while you dial in the scale.