FAL Flux 2 LoRA Edit

FAL Flux 2 LoRA Edit is the API wrapper for fal-ai/flux-2/lora/edit - Flux 2 editing with a LoRA bolted on for style control. If you've ever wanted to run an edit through Flux 2 with your own trained style or character LoRA but didn't want to host the 56B model stack, this is the compromise: fal runs the model, you supply the reference images, and the lora_url input is where your style lives.

The mechanism is a straightforward API submission. You send up to 3 reference images (image1 required, image2/image3 optional - Flux 2's multi-reference capability is its one genuinely unbeatable local feature, and this node exposes a slice of it), a text prompt describing the edit, and optionally a LoRA. The endpoint's lora suffix is doing real work: it's the fal endpoint variant that accepts lora_url and lora_scale and applies them to the edit pipeline. So the model gets the edit instruction, the reference for what to keep, and the LoRA for how it should look.

The inputs that matter:

• api_key - FAL key (node input or FAL_KEY env var).
• prompt + image1 - the edit and the source. Default prompt is the same "Make this donut realistic" sanity check used across the FAL edit nodes.
• lora_url - a hosted LoRA file URL. This is the differentiator. Drop a URL to a LoRA you trained or downloaded; leave blank to run vanilla.
• lora_scale (default 1.0, 0–2) - how hard the LoRA bites. Same intuition as local ComfyUI: 0.8–1.2 is the sweet spot for style, above that and the LoRA starts to fight the prompt.
• num_inference_steps (default 28), guidance_scale (default 2.5), enable_prompt_expansion (off by default - a fal-side rewriter that prettifies your prompt; the author of the LoRA node gives you the choice, and it's genuinely useful when your prompt is short and you want the model to fill in), acceleration (none/regular/high).
• width/height, seed, output_format, enable_safety_checker - the standard tail.

Outputs: image, image_path, generation_info, generation_time, saved to output/API/FAL/flux-2-lora/ with metadata.

Where this fits: Flux 2's multi-reference editing is the capability the KB says "nothing else matches locally," and putting a LoRA on top gives you a repeatable style. The 2026 edit-model picture makes the tradeoff concrete - Klein 9B and Qwen-Image-Edit are the serious open editors, and Flux 2's full stack is heavy enough that most people only touch it via an API. This node is exactly that on-ramp: the parts you'd normally spend a weekend downloading (the 32B transformer plus the Mistral-3 24B encoder) are handled for you.

Honest caveats. It's paid, per call, and LoRA + multi-reference + high steps is the most expensive combination in the pack. And there's a subtle mismatch worth knowing: the LoRA URL is a hosted file - the workflow assumes the LoRA is on fal's infrastructure (you can often point it at a public file URL), not sitting in your models/loras folder like you're used to. If your LoRA is local-only, this node isn't the path; you'd want to host it or run the model locally.

Install is the pack standard: Manager search "Node Fun", or clone https://github.com/nofunstudio/Node_Fun_ComfyUI into custom_nodes, pip install -r requirements.txt, restart. Given how rarely this node is Googled - it's one of the least-viewed in the pack - assume you're early: the endpoint may lag fal's newest Flux 2 variants, so if a run errors, check whether fal-ai/flux-2/lora/edit still exists before blaming your key.