arkennemasis fal · Topaz Upscale Video Precision · $0.01-0.06/s

Every generative upscaler in this ecosystem has the same failure mode: it invents. Detail that wasn't there, faces that are almost right, texture where the source had mush. Sometimes that's great - you're rescuing a render you can't re-render. When the source is real footage, it's a problem. Topaz's precision models are the other approach: non-generative enhancement, up to 4x, staying faithful to what was actually filmed.

That's the whole pitch, and the price is small enough to make it a default: $0.10 per 10 seconds at 720p, $0.20 at 1080p, $0.60 at 4K for the standard precision models. A one-minute clip upscaled to 1080p costs about eighty cents. For comparison, Topaz is the paid tool the open-source crowd benchmarks against constantly, and this node is that tool as a per-use call instead of a subscription.

Picking a model

The model dropdown is where the actual decision lives, and the tooltip is a decent map: Proteus fits most real-world footage, Proteus Natural is a softer variant (2x only), Iris recovers faces, Dione deinterlaces legacy or interlaced sources, Artemis denoises and sharpens degraded footage, Gaia refines high-quality footage and CG, Gaia 2 upscales animation at 2x, Rhea goes after maximum fine detail with an internal 4x pass, and Theia gives you manual detail/fidelity control.

For archive work - VHS rips, old camcorder footage, anything with combing - Dione is the one people forget exists. For animation, Gaia 2. For a modern render, Proteus and leave it.

The settings

upscale_factor runs 1 to 4 and doubles width and height at 2.0. target_fps is -1 by default, which keeps the source rate; set it to something else and frame interpolation turns on, which is billed as more output and is a bigger deal than it sounds - interpolated frames are generated, so that's where a faithful upscale starts adding invented content.

Then a row of five knobs that all default to -1, meaning "leave it to the model": compression (artifact removal), noise (noise reduction), halo (halo reduction), grain (film grain, 0–0.1), and recover_detail (higher keeps more of the original). You mostly don't touch these. If an upscale comes back looking plastic, drop noise a notch; if it's smearing fine texture, raise recover_detail.

H264_output defaults to off, which means H265. Turn it on if the video is going somewhere that won't play H265 - this is the most common "the file is fine but my editor/phone/browser hates it" moment in this pack.

Outputs: video (VIDEO, wire it to a save or the pack's assembly nodes), video_path, and info with the request id, cost estimate, saved files and fal's full answer. Results land in output/fal/.

Pack plumbing: max_cost_usd (default $20), reuse_identical_run, extra_json, max_concurrent (1 = one paid fal call at a time across the run).

Install

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

Manager → arkennemasis. Then:

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

in the .env next to run_nvidia_gpu.bat, then restart. The node appears under arkennemasis/fal/Video/Topaz.

Worth knowing

• Cost scales with output resolution, duration and frame rate, not input size - so a 4x upscale of a 240p clip to 4K is priced like any other 4K job. Set max_cost_usd low while you're experimenting.
• Frame interpolation is a content change, not a sharpening pass. It's the setting most likely to make "precision" not apply.
• reuse_identical_run is on by default. Changing model or upscale_factor changes the inputs and triggers a real run; changing something downstream doesn't, and you'll get the old file back.
• If you have a local GPU and the source is already decent, the open alternatives (SeedVR2, NVIDIA's RTX 2-pass) are worth a look before you pay per clip - Topaz wins on convenience and on legacy-footage presets, not on being the only option.