arkennemasis fal · Flux Subject · $0.04/MP

Flux Subject puts a specific thing back into a generated scene. Not a face, in the PuLID sense - a product, a bag, a mug, a brand identity, a style reference. You supply the picture of the subject, describe the scene you want it in, and it renders that scene with your object in it. Schema-wise it's Schnell with a subject input: fast endpoint, 8 default steps, and the cheapest subject-conditioning option in this pack at $0.04/MP.

That price point is why it's interesting. Four cents a megapixel is a hair above PuLID, well under Fill, and far under the Ultra endpoints, so you can generate a product-shot matrix - five backgrounds, four angles - without doing arithmetic first. If you've done the customer-photo-to-catalogue thing with local LoRAs, this is the "no training, no dataset, just a reference and a prompt" version of the job, and the trade you're making is fidelity for speed. Don't expect a LoRA's precision from one reference image.

Inputs

• image - the subject. One image, clean crop, plain-ish background if you have the choice.
• prompt - the scene you want built around it. The description does the heavy lifting; the subject socket only holds one thing constant.
• image_size - square_hd is the default here (unlike the landscape default on the other Flux nodes in this pack), with square, the landscape/portrait presets and custom plus image_size_width/image_size_height available. Billing is per output megapixel, so this is the price dial.
• num_inference_steps - default 8, up to 50. The endpoint is the "super fast" Schnell branch; the defaults are where the speed is.
• guidance_scale - default 3.5.
• num_images (up to 4), seed, output_format (png by default here), enable_safety_checker.
• The shared four: max_cost_usd, reuse_identical_run, extra_json, max_concurrent.

Outputs

images and image_paths (files under output/fal/fal-ai_flux-subject/), seed as an INT, plus prompt, timings, has_nsfw_concepts and info - JSON containing the request id, the cost estimate, the saved files and fal's full answer. Wiring is the usual: image socket into SaveImage, or into a fill/erase node, or into an upscaler if 1 MP isn't enough.

Install

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

or install comfyui-arkennemasis from ComfyUI Manager. Restart, then add the key line to the .env in the folder that holds run_nvidia_gpu.bat:

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

No downloads, no Python dependencies for this node, no key widget - the pack reads the .env and nowhere else, which is exactly why your key can't end up inside a workflow file.

Where it goes wrong

Subject conditioning is sensitive to the reference crop. A tight, well-lit, single-object photo with the whole object in frame works; a cluttered shot with three similar items gets you an amalgam of all three. If the output keeps morphing your product, that's the model's guess showing - crop harder and describe less.

There's also an expectation mismatch worth naming: "Subject" sounds like it guarantees consistency across a set, and it doesn't - it biases one generation toward one reference. For a run of twenty images that must all show the same object, you want a trained LoRA, or the pack's Variation pipeline, or an edit model. This node is the fast first draft of that project, not the guarantee.

And the pack-wide caveats: the safety checker returns black images rather than errors, and switching it off needs account authorisation on fal; reuse_identical_run hands back cached files for identical inputs unless you turn it off; a missing key or an empty fal balance stops the run before it bills; and the free Recover Result node rescues a job whose ComfyUI died mid-wait.