arkennemasis fal · Flux 2 LoRA Gallery Virtual Tryon · $0.021/MP

Two pictures in, one out: a person and a garment, and you get the person wearing the garment. That's the whole node. fal's flux-2-lora-gallery/virtual-tryon endpoint runs at $0.021 per processed megapixel, and the pack exposes it with the LoRA strength as a knob - which is the single dial that decides whether the clothes fit the body properly or float slightly above it.

How it works

Flux 2 is unusually good at multi-reference conditioning - the KB treats that as its signature strength, and it's the reason the flock-on-4B try-on demos that circulated worked at all. fal trained a LoRA on top for the try-on job specifically, and the node uploads both pictures to fal's CDN before submitting the request. Results go to output/fal/fal-ai_flux-2-lora-gallery_virtual-tryon/.

Inputs

image_1 is required and the tooltip is blunt about the order of business: person image and clothing image. Sockets run up to image_6, and batches are sent in socket order (first picture of socket 1 is number 1) - which matters here because the second image you send is the garment.

prompt describes the result. Empty is allowed and works, but a short instruction about pose and framing ("full-body shot, standing, neutral studio light") helps the model decide what to render rather than cropping to the torso.

lora_scale (0–2, default 1) is effect strength. Under 1 leaves the original clothing partly visible; over 1 makes the garment dominate, sometimes at the cost of body shape. If the garment is landing badly, this is the first thing to move.

guidance_scale (0–20, default 2.5) and num_inference_steps (4–50, default 40) are the Flux 2 settings - 40 steps is not cheap, so halving it is a reasonable exploration mode. acceleration is none or regular. image_size defaults to square_hd; full-body try-on usually wants portrait_4_3, or custom with the width and height set. num_images (1–4) generates variations in one paid call, which is how you'd judge a garment, and seed at -1 randomises. output_format defaults to png; the safety checker returns flagged output as black images.

The shared options: max_cost_usd (default $20, 0 disables) stops a run before anything is uploaded when the estimate is too high; reuse_identical_run returns an identical already-paid result instead of charging again; extra_json merges raw JSON into the request for fal fields with no box; max_concurrent (default 1) limits how many paid fal calls run at the same time across the graph.

Outputs

images is the result - a normal IMAGE batch for saving, compositing on a background, or dropping onto a contact sheet. image_paths, seed, prompt and info come with it; info is the JSON holding the request id, the estimate and fal's full answer.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Restart, then in the .env next to run_nvidia_gpu.bat:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

The fal provider pulls in no Python packages, so there's nothing to install for these nodes.

Where people get burned

Garment fidelity is the honest weak point: patterns, fine text and logos get reinterpreted, because the LoRA is re-rendering the clothing from an image rather than texturising a 3D body. Flat-lay product photos with clear edges do better than a garment photographed on a hanger in bad light. Second, the person's proportions and skin tone can shift a little - acceptable for a lookbook, worth checking if it's a real customer's photo and they'll notice.

Third, cost: it bills processed megapixels including both inputs, so a 2MP person plus a 1MP garment plus a 1MP output is around $0.084, and a four-image variation pass roughly quadruples it. And the deal you're making is the same as everywhere in this pack: both photographs are uploaded to fal before anything renders, so don't put a client's face or an unreleased design through it without deciding that's fine first.