arkennemasis fal · Flux 2 LoRA Gallery Add Background · $0.021/MP

Every product shoot ends the same way: a clean cut-out of the object on white, and a client who wants it in a café. This node is the shortest path between those two things - you feed it the white-background shot, tell it what background you want, and it hands back a composited image at $0.021 per processed megapixel.

What it is

One of eleven called-out style endpoints fal publishes under flux-2-lora-gallery: a fixed LoRA trained for one job, bolted onto Flux 2, exposed here as a single node. The pack is a thin client - it uploads your picture, submits to fal's queue, waits, downloads into output/fal/fal-ai_flux-2-lora-gallery_add-background/. There's a lora_scale box because the style is applied as a LoRA, and that's the single most useful knob on the node.

Inputs

image_1 is required and the tooltip is specific about what it wants: an image with a white or clean background. The model is adding, not masking - feed it a cluttered photo and you get a cluttered photo with more clutter. Sockets run up to image_6 and batches count, so you can hand it several product shots in one request - they're sent in socket order, first picture of socket 1 is number 1, and every picture is billed as processed megapixels.

prompt must start with Add Background followed by your description - the default is literally Add Background forest. The tooltip says to keep the prefix, and the model expects it. This is the part where you get to be specific: "Add Background a sunlit café terrace, shallow depth of field, no visible text".

lora_scale (0–2, default 1) is the strength of the effect. Push past 1 when the model is being timid about the new scene; pull it under 1 when it starts eating the product.

guidance_scale (default 2.5) and num_inference_steps (default 40) are the generation dials - 40 steps is generous, and you can trade some of that for time. acceleration is none or regular. image_size defaults to square_hd, so a vertical product shot comes back square unless you pick a preset or custom; num_images (1–4) multiplies the bill; seed at -1 lets the model choose; output_format and the safety checker are the usual pair - flagged images return black, and disabling the checker needs fal account authorisation.

max_cost_usd stops the run before upload if the estimate is over the cap. reuse_identical_run fingerprints your settings and the picture bytes so an identical re-run is free. extra_json passes through fal fields the node has no box for, and max_concurrent keeps simultaneous paid calls in check.

Outputs

images is the composited result - straight into a save node or a contact sheet. image_paths, seed, prompt and info are the record: info holds the request id, the cost estimate, the saved files and fal's full answer. That JSON is worth keeping when you're billing someone for a batch.

Install

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

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

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

Nothing to pip install for the fal nodes - standard library only. There's no key field on the node, by design, so a key never ends up inside a shared workflow.

Where people get burned

Contrast and shadow. A white-background product shot has flat lighting; drop it into a forest and the object often still reads as flat, because the model is compositing rather than relighting. Ask for the light in the prompt ("warm late-afternoon sun from the right") or expect to match it downstream. Second: the model loves adding set dressing near your product - text, signage, small props - so say what you don't want.

Third, cost: it bills processed megapixels, which includes the input, so a 2MP source and a 1MP output is closer to $0.063 than the $0.021 the badge suggests. And this is a paid third-party call - your product photo is uploaded to fal's CDN before anything renders. For a client's unreleased product, check that's acceptable before you batch forty of them.