Topaz Labs Wonder 3.5

Topaz is the software that local upscaling threads keep benchmarking against - the paid default the open-source nodes are trying to displace. This node is Topaz's Wonder 3.5, not running in Topaz's desktop app on your machine but as a per-image cloud call through Runware's ComfyUI pack. You get the commercial model in your graph with no Topaz subscription and no GPU that can feed it. The trade is that every upscale is metered, so it's the occasional "this one image is worth it" tool, not your daily upscaler.

Understand what Wonder actually is before you wire it in. Per Runware's model page, it's a generative image enhancement and upscaling model - the "more detail" job in the upscaling taxonomy, not the "more pixels" job. It recovers texture and realism from soft, compressed, or degraded sources and, usefully, handles structured content like tables, charts, and text-heavy graphics without melting them into mush. If your source is already clean and you just want it bigger, an ESRGAN model is free, instant, and the right answer; this is for restoration and repair where you want the model to invent plausible detail.

What you set. The only required input is image, and you feed it any IMAGE in your graph. That's the nice part - unlike the pack's video-side Topaz node, here you just wire in the output of your last sampler. Three controls actually matter:

• upscaleFactor (2–6, default 2) - the multiplier, and it quietly caps your input size: the tooltip notes output can't exceed 256 megapixels, so a big source at 6x will be rejected.
• settings.enhancementStrength - low / medium / high, defaulting to high. If the result looks overcooked or plastic-y, this is the first knob to turn down, not a mystery detail slider.
• outputFormat / outputQuality - PNG if you're going to keep editing, JPG at the default 95 otherwise.

Everything else (TTL, an advanced_json escape hatch for settings.grain) you can ignore. Output is a native IMAGE under the name image, so it plugs straight into Save Image or downstream nodes like any upscaler result.

Install. This ships in the Runware pack, so installing it is the same as any other Runware model node:

# ComfyUI Manager (recommended): Custom Nodes Manager → search "Runware" → install → restart
cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware   # manual alternative
pip install -r ComfyUI-Runware/requirements.txt

Restart, then add a Runware API key (ComfyUI Settings → Runware API key, or the RUNWARE_API_KEY environment variable). No model files to download - Wonder runs on Runware's side.

Gotchas. First, the obvious one: it's a cloud call, so your image leaves the machine and each queue has a price shown in the node's title bar. Second, "generative upscaler rewrites faces" applies here as much as it does to SUPIR - if identity matters, keep Wonder's job to background and texture repair, or expect the subject to drift. And don't confuse this node with the pack's video Topaz node (Proteus); Wonder is the image one, and it's far easier to drop into an existing workflow because it takes a plain IMAGE.