arkennemasis fal · Flux 2 LoRA Gallery HDR Style · $0.021/MP

HDR here does not mean the bracketed-exposure technique photographers argue about. It means the cranked, hyper-detailed, faintly surreal look that flooded stock photography in the 2010s and never quite left - halos around every edge, colours with the saturation slider at 11. fal sells it as a LoRA on Flux 2 for $0.021 per megapixel, and the pack wraps it in one node.

What it is

Text-to-image, style-only, no image input. The endpoint is flux-2-lora-gallery/hdr-style; the node submits your prompt, waits on fal's queue, and drops the results into output/fal/fal-ai_flux-2-lora-gallery_hdr-style/. There's one quirk worth flagging because it's unusual: per the tooltip, the trigger word Hyp3rRe4list1c is prepended automatically. You don't type it, and typing it yourself probably doesn't hurt but isn't required. Every other style node in this gallery asks you to remember its trigger token.

Inputs

prompt - the subject and framing, written as sentences. Flux 2's encoder is a language model; masterpiece, 8k, ultra detailed is noise to it. lora_scale (0–2, default 1) is the look's intensity, and it's the only control that changes the character of the image rather than the content: past 1 the effect gets cartoonish, below about 0.6 you're paying for a LoRA that isn't doing anything.

guidance_scale (0–20) defaults to 2.5; num_inference_steps runs 4–50 and defaults to 40, which is generous for a style pass - 20 is a fair first test. acceleration is none or regular. image_size defaults to landscape_4_3 with the usual presets and custom; num_images (1–4) batches in one paid call, which is how you'd use a style node - four takes, pick the least crunchy one. seed at -1 lets the model choose, output_format defaults to png, and the safety checker returns flagged output as black images.

Shared options on every fal node in this pack: max_cost_usd (default $20, 0 disables) blocks a run before anything is uploaded if the estimate is over; reuse_identical_run returns an identical already-paid result instead of billing again; extra_json merges raw JSON for fal fields with no box on the node; max_concurrent (default 1) stops a big canvas from firing simultaneous paid calls.

Outputs

images, image_paths, seed, prompt, info. The info JSON holds the request id, the cost estimate, the saved files and fal's full answer - the only honest record of what a session cost, since estimates and badges can't see things like LoRA size surcharges.

Install

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

Restart, and put the key in the .env next to run_nvidia_gpu.bat:

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

No pip install is needed for the fal nodes - that provider uses Python's standard library.

Where people get burned

The honest problem with this look is that it is currently unfashionable. Photoreal work in 2026 is chasing imperfection - smartphone-snapshot realism, visible sensor noise, skin that isn't airbrushed - and HDR is the opposite instinct, so if your goal is a believable photo of a person, this is the wrong node and you'll spend the evening dialling lora_scale down toward zero. Where it earns its keep is deliberately stylised output: landscapes, sci-fi vistas, game art, anything where "hyper-real" reads as a choice rather than a mistake.

Second, letters and faces: Flux 2 will happily put garbled text and slightly wrong hands into a high-contrast scene, and HDR's edge halos make both more obvious. Third, cost - $0.021 a megapixel sounds trivial until you've done sixty iterations across a session, and every one of them was a billed upload to a third party.