Upscale Pro

The name is a lie in exactly one way

"Upscale Pro" sounds like another local model to add to your ever-growing models/upscale_models folder. It isn't. No model file, no VRAM spend, no GGUF to fuss over. This node takes a rasterized logo, sends it to Augment Studio's servers, and gets back a sharp 2x version. It's built for one job: taking logos, icons, and flat artwork and making them print-ready - lines crisp, colors clean - before you vectorize.

The catch up front: it's a paid cloud service, not a free local one. You need an API key from augmentstudio.app/pricing, each run costs a credit (the node draws a small "1 export / run" badge on itself), and your image leaves your machine. That last one matters more than the cost if you upscale client work under an NDA.

Why it exists in the upscaling ladder

Upscalers fall into three jobs - more pixels, more detail, more pixels over time. This one is squarely job one, aimed at a very specific source. It's not trying to invent pores and eyelashes like SeedVR2, and it won't repair a damaged photo like SUPIR. Throw a portrait at it and you'll be disappointed. It's for flat artwork with hard edges, where a generic upscaler smears the boundary between a shape and its background. The pack's own pitch calls it "a professional alternative to Magnific AI" - same cloud-upscale-for-creators market, built straight into ComfyUI.

So: generate your logo or icon, drop this in, get a clean 2x raster, and send that to your vectorizer. One step, not a five-node chain.

How it actually works

The code is a thin HTTP client, and that's the whole story. When you run it:

1. It POSTs your image to https://augmentstudio.app/api/process with your key as a Bearer token and scale=2.
2. It polls a job status endpoint every second, up to a 120-second ceiling.
3. When the job's done it fetches the result, converts it to an IMAGE tensor, and hands it back.

No local inference at all. The only dependencies are requests, numpy, torch, and Pillow - all already sitting in a normal ComfyUI install, so nothing heavy to download and no model files to hunt. Your GPU idles while their servers work, but the flip side is it needs internet and their API to be up. The 120-second timeout is the practical gotcha: a slow queue or a hiccup and you get "Timed out after 120s."

The inputs that actually matter

The two inputs are obvious:

• api_key (required, string) - paste the key from the pricing page. Nothing happens without it.
• image (optional) or image_file (optional) - you need one or the other. Feed image from your generation, or pick a file from ComfyUI's input folder. Neither connected and the node throws "No image provided" rather than guessing.

scale is technically an input but it's a dropdown with exactly one choice right now: 2x. Don't go hunting for a 4x mode; it doesn't exist yet. There's also a trigger input if you're chaining steps with the pack's flow-control nodes.

The outputs are image (your upscaled result, wire it anywhere an image goes), trigger (for chaining), and json_result, which carries metadata like output width, height, and scale. Fun detail: the pack's own UI script hides the json_result output on screen, so don't go hunting for it in the node editor.

Installing it

ComfyUI Manager is the easy path - search "augment Upscale Pro" and install. Manual install works too, and here's a real trap: the pack's own README gives a clone URL that 404s. The repo that actually exists is:

cd ComfyUI/custom_nodes
git clone https://github.com/augment-lib/upscale-pro

Then restart ComfyUI. No models, no extra config - just the key.

To avoid pasting your key into every workflow, the README suggests the companion augment-ComfyUI pack: set the key once with Variable Set (String), wire its text output into api_key, and future workflows pull it from a Variable Get (String).

What to expect

It's new - the repo's first commits are from March 2026, there's essentially zero community discussion of it anywhere, and no GSC footprint yet. Not a knock - just don't expect a long history of workflows to crib from. For the narrow job it does - logos and flat art to print-ready raster - the appeal is one node, zero VRAM, zero local setup, and a small per-run fee to skip the fiddly parts. If your source is photos, remember the three-jobs rule and reach for SeedVR2 or an ESRGAN model instead. If it's a logo you need to stop being blurry, this genuinely might be the fastest path.