Topaz Photo Upscaler (Autopilot) JNK

Topaz Photo AI is the paid benchmark people keep comparing everything else against - the KB's upscaling panel calls it "the paid default" at 200+ Reddit threads a year, and the honest community take is that it's quick, easy, and quite good, though a well-built ComfyUI workflow can beat it on quality. If you already pay for Topaz and want its Autopilot mode inside your graph instead of a separate app, this node is the bridge. "Topaz Photo Upscaler (Autopilot) JNK" takes your image tensors, hands them to the installed Topaz Photo AI executable, and brings the upscaled results back into ComfyUI as tensors - with a side of autopilot settings for the curious.

The mechanism is worth knowing, because it's not doing the upscaling itself. For each image in your batch, the node writes a PNG to ComfyUI's temp folder, then shells out to tpai.exe with a command like:

tpai.exe --overwrite --output <temp/upscaled> --compression 2 --format png --showSettings "<input.png>"

Topaz does its Autopilot thing and writes the result to the output folder; the node loads it back and returns a tensor. The --showSettings flag makes Topaz print its final decision as a JSON block in stdout, which the node parses ("Final Settings for {...}") and pops the autoPilotSettings object out of. That object is what comes back as the first output - a nicely formatted string showing exactly what model/settings Topaz picked, so you can see what "autopilot" decided without opening the app. It's genuinely useful when you're tuning: you can see whether Topaz chose Gigapixel or Photo, what it did with noise, and adjust its preferences accordingly.

Because Autopilot is the whole game here, the node can't control the actual upscaling quality - that's configured inside Topaz Photo AI. The source code's comment spells out the setup: open Topaz Photo AI → Edit → Preferences → AutoPilot Settings, and configure what Autopilot does there. The node's job is just to trigger it. If you want custom per-image settings, the node author's answer is to configure Autopilot in Topaz's preferences rather than fight for granular control here.

The inputs that matter:

• images - the batch to upscale. Each one is a separate tpai.exe call, so a 10-frame batch means 10 launches.
• compression - 0–10, controls PNG compression on the upscaled output (default 2). Lower = bigger files, higher = smaller.
• tpai_exe - path to the executable. Defaults to the standard Windows install (C:\Program Files\Topaz Labs LLC\Topaz Photo AI\tpai.exe). The README and tooltip describe it as "the folder containing tpai.exe," but the code checks the path directly - point it at the executable file itself and you'll have a better time.

Both outputs are lists, one entry per input image: autopilot_settings (the parsed settings strings) and IMAGE (the upscaled tensors).

Install: ComfyUI Manager → search "JNK" → Install, then restart. Or:

cd ComfyUI/custom_nodes/
git clone https://github.com/Aljnk/ComfyUI-JNK-Tiny-Nodes.git

Then restart. The real dependency isn't Python - it's a licensed copy of Topaz Photo AI installed on the machine, and it's effectively Windows-only (the default path is a Windows path, and it shells out to a .exe).

Gotchas: this is the pack's most fragile node, in the best-informed sense. No Topaz install, or a wrong tpai_exe, and it raises a ValueError ("Topaz AI Upscaler not found") - it does not silently fall back to a normal upscale. If parsing the settings JSON fails (Topaz output format changed, which happens), it prints a ---JNK---> TopazPhotoAI_JNK ---> Error to the console and returns {} for settings rather than crashing - but the upscale itself still runs. Speed: launching a GUI-backed binary per image is not fast, so don't feed it long video batches. And remember the quality ceiling: you're paying for Autopilot's defaults, and the community's consistent finding is that a tuned tile-ControlNet workflow can outdo Topaz on demanding sources - use this node when you want Topaz's convenience in the loop, not when you're chasing the absolute best result. For that, the KB's advice still holds: a free in-graph upscaler will often beat it.