PixelPanda: AI Upscale

PixelPanda: AI Upscale is the pack's workhorse, and the node the README's print-on-demand pipeline is built around: generate art → upscale 4x → remove the background → upload to your store. It's also the node in ComfyUI-PixelPanda with actual knobs to turn, which is refreshing in a pack where everything else is one-button. Drop in an image, pick a scale, pick a quality level, and the PixelPanda API's /upscale endpoint sends back a bigger image. 1 credit at 2x and 4x, 2 credits at 8x.

What kind of upscaler is this, really

Worth getting straight before you pay for it. The KB's upscaling doc splits the field into "more pixels" and "more detail", and this node is firmly the second kind: a cloud AI upscaler that invents plausible detail rather than just resampling. That's what makes it useful for taking a 1024px render to print size with actual texture instead of mush - and also what makes it the wrong tool if your source is already sharp and you just need it bigger, because a free ESRGAN model does that job instantly and can't hallucinate. The generative caveat from the local world applies here too: every generative upscale rung rewrites fine detail, so don't point it at a face you need to stay recognizable and expect a copy.

One difference from running SeedVR2 or SUPIR locally: you can't tune it. The API gives you three quality presets, not a denoise value, and you won't know if "high" bought you anything until you pay for a second run at 8x - the 8x tier costs 2 credits precisely because it's the expensive generation.

The inputs that matter

• image - what you're upscaling.
• api_key - pk_live_… key, or blank to use the PIXELPANDA_API_KEY environment variable.
• scale - 2, 4, or 8. Default 4. This is the whole game: 2x for light cleanup, 4x for the standard print-on-demand jump, 8x when you genuinely need a giant canvas. Remember 8x is 2 credits.
• quality - fast, balanced, or high, default balanced. The speed-versus-fidelity trade. Start at balanced; only reach for high when you can see artifacts at balanced and don't mind waiting longer (the node will wait up to 300 seconds regardless).

One IMAGE comes out, at your chosen scale, ready for the next node.

Install and key

cd ComfyUI/custom_nodes
git clone https://github.com/RyanKramer/ComfyUI-PixelPanda

Restart ComfyUI and it's under image/PixelPanda; or install via ComfyUI Manager by searching "PixelPanda". No extra dependencies - the pack ships with an empty requirements list and uses only torch/numpy/Pillow already in ComfyUI. Grab a key at https://pixelpanda.ai/profile; it goes in the node field or the PIXELPANDA_API_KEY env var. The $1.99 starter pack needs no subscription and credits never expire.

When to pay vs. when to run local

This is where you should be honest with yourself. The upscaling KB is clear that the "just give me pixels" tier got cheaper, not more expensive - Lanczos and 4x ESRGAN models are free, instant, and run on a potato, and for an already-sharp source they're the correct answer. If you've got the VRAM and upscale constantly, a local generative model is free after the electricity and gives you control this node can't. PixelPanda makes sense for the same reasons the rest of the pack do: no spare VRAM, no desire to maintain another model, an occasional batch that needs print resolution, or a workflow where one metered node beats downloading weights. A 1-credit charge for 4x is a fine impulse purchase; building a high-volume pipeline on it is a recurring cost you should price before committing.

Gotchas

Only the first frame of any batch is processed - no frame-by-frame upscaling of video. The call needs the network and genuinely can take a while at high quality; it's not frozen, it's waiting, and the console prints [PixelPanda] /upscale ok - credits used/remaining when done. Empty credits read as a clean 402 error, a bad key as 401. And the pack-wide caveat: PixelPanda is a small, new service with almost no independent community footprint, so verify it's still alive before you make it a permanent stage of a production workflow.