arkennemasis fal · Flux General · $0.075/MP

Every other Flux endpoint on fal is a specialist: this one is the general-purpose endpoint, and it's general in the way a Swiss army knife is. LoRA, ControlNet, control LoRAs, IP-Adapter, EasyControl, reference-only conditioning, a scheduler choice, negative prompts via NAG - all present, all optional, all in one node. It's the most configurable thing in this pack and the one most likely to be overkill.

It's also the most expensive per megapixel in the Flux line at $0.075/MP - three times Flux Dev - so the question isn't "which is better" but "do I actually need one of these extensions".

When you reach for it

You need a LoRA you don't have locally. You need two conditions at once (a pose plus a depth map). You want reference-only conditioning, which is somewhere between img2img and IP-Adapter. Or you want negative prompting on Flux without the local NAG install.

Otherwise, use the specialist node. flux-dev at a third the price is the same base model.

The JSON fields, and what they actually want

Six inputs on this node take JSON, and they're the reason the node looks intimidating:

• loras - the extra LoRAs to merge in.
• control_loras - LoRAs that consume a control image.
• controlnets - note the tooltip: only one is supported at the moment.
• controlnet_unions - same single-controlnet caveat.
• ip_adapters - IP-Adapter entries.
• easycontrols - EasyControl inputs.

Empty means "leave it out", and empty is the right value for all six until you have a specific reason. Here's the practical gotcha nobody mentions: these expect URLs to files fal can fetch. See a LoRA on CivitAI you want to try? Its safetensors has to live somewhere fal can download it from - a Hugging Face repo, your own storage, an S3 bucket. LoRAs sitting in your ComfyUI models/loras folder are invisible to this endpoint. That one fact kills the "just point it at my local library" expectation instantly.

The NAG knobs, which are the real reason to look twice

Flux is guidance-distilled, and on a guidance-distilled model negative prompts are inert rather than weak - there's no unconditional pass to steer at guidance 1. That's why "blurry, ugly, deformed" in a negative field does nothing on modern Flux. The workaround the community settled on is NAG (Normalized Attention Guidance), which intervenes in attention space instead of the guidance term, and the author's own caveat is not to exceed nag_scale 3.

This endpoint exposes it, and it's the only node in this pack where negative_prompt is a serious field:

• negative_prompt - steer away from unwanted features. Empty leaves it out.
• nag_scale (default 3) - higher pushes further from the negative. Do not exceed 3.
• nag_tau (2.5) - normalisation of the hidden state.
• nag_alpha (0.25) - final weighting of the positive/negative direction.
• nag_end (0.25) - the proportion of steps NAG applies for; after that the original attention processors take over.

If you've ever wondered whether negatives do anything on Flux without a NAG install: this node is the cleanest place to find out.

The CFG fields, and the expensive one

use_real_cfg (default false) restores classical SD1.5-style CFG. The tooltip is explicit that it increases generation times and price. real_cfg_scale (3.5) is that guidance value. use_cfg_zero (false) applies CFG-Zero-star initialisation, with the arXiv reference right there in the tooltip. guidance_scale (3.5) is the normal Flux guidance.

Leave all three alone unless you know what you're buying: use_real_cfg literally inflates the bill.

Reference-only, and the timestep window

reference_image plus reference_strength (0.65), reference_start (0.0) and reference_end (1.0) is reference-only generation - neither img2img nor full IP-Adapter, and the start/end window is the same idea as ControlNet's "release the condition once composition has formed". There are also six optional fill_image_* sockets for Redux-style fill prompting.

Then the sampling controls: scheduler (euler default), base_shift (0.5), max_shift (1.15) for the timestep schedule, and use_beta_schedule (false). num_images is capped at 1 - the tooltip explains it's fixed for streaming output - and output_format defaults to png here, not jpeg.

Outputs and install

Outputs are the standard set: images, image_paths (in output/fal/fal-ai_flux-general/), timings, has_nsfw_concepts, prompt, seed, and info with the request id, cost estimate, files and fal's full answer.

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

Restart ComfyUI, or use ComfyUI-Manager → Install via Git URL. Only httpx is declared, for the pack's Codex nodes; the fal nodes use the standard library. No weights touch your disk.

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

in the .env next to run_nvidia_gpu.bat (portable) or in your ComfyUI folder. Never a widget - deliberately, so the key can't end up inside an exported workflow.

Money, and the trap

At $0.075/MP a 1024×1024 is ~7.5¢ and a 2048×2048 is ~30¢, and use_real_cfg pushes it higher for the same picture. max_cost_usd (default $20, 0 = no cap) refuses before upload when the estimate is over. reuse_identical_run defaults on, so an unchanged graph returns the file you already paid for - which, on a node this expensive, is a feature you'll appreciate.

The real trap is subtler: a typo in any of those JSON fields fails the run before anything is sent, which is good, but the error comes back as "is not valid JSON" and tells you nothing about what the field expected. Paste the JSON into anything that validates it first.