Load Image → MegaPotato Upscale (API) at ×8 → Save Image: 648 × 432 in, 5 184 × 3 456 out, about 35 seconds on a warm server, 2 chips. The first job after a quiet period waits 2–5 minutes for a GPU to start; the node says queued meanwhile.
Photo: Jimmy Chan / Pexels
MegaPotato Upscale (API) sends an IMAGE to the MegaPotato AI upscaler (megapotato.app) and returns it enlarged ×1, ×2, ×4 or ×8 as an IMAGE. Nothing runs on your GPU and there is no model to download: the photo goes to MegaPotato's cloud GPUs, and the rest of the workflow does not know the difference. It is a photo upscaler, not a creative one: it sharpens what is there and keeps the picture close to the original. In the cover picture a general-purpose image model redrew the woman's face; MegaPotato kept it.
In our October 2026 benchmark of 18 AI upscalers it came first on both tests and was the cheapest of the four leaders; the whole method, every photo and every score are there.
Install#
ComfyUI Manager: search for MegaPotato. Or by hand:
cd ComfyUI/custom_nodes git clone https://github.com/MegaPotato-AI/comfyui-megapotato
No extra Python packages. Restart ComfyUI and the node is in the menu under image/upscaling; an example workflow comes with it.
API key#
Create a key at megapotato.app/account and put it in an environment variable before starting ComfyUI:
export MEGAPOTATO_API_KEY=mp_live_... # Linux / macOS set MEGAPOTATO_API_KEY=mp_live_... # Windows cmd $env:MEGAPOTATO_API_KEY="mp_live_..." # Windows PowerShell
Or save it in custom_nodes/comfyui-megapotato/megapotato_api_key.txt (git-ignored). A test key (mp_test_…) runs the whole flow and charges nothing: every job comes back with a sample image. Live keys unlock with your first chip pack.
The node#
| Input | What it does |
|---|---|
image | The photo or photos. A batch becomes one job per image, up to 4 in parallel. |
upscale | 1 enhances at the same size; 2, 4 and 8 enlarge. |
api_key_override | Only if the environment variable is not an option. |
timeout_minutes | How long to wait, 15 by default. |
rerun | The same photo at the same factor is never charged twice. Change this number to force a fresh run. |
Outputs: image, job_id (also in your history on the site) and cost_chips. Input up to 6 000 px per side and 20 MP, output up to 16 400 px per side, so ×8 takes inputs up to 2 050 px and ×4 up to 4 100 px; the node checks this before uploading. A ×8 of a 2 000 px photo is about 3 GB of RAM inside ComfyUI: save it straight away.
Cost#
Priced by the size of the photo and the factor, the same as on the site, and printed before the upload. Chips are held when the job is submitted and refunded if it fails. If a run times out, the job keeps going on the server: run the node again with the same photo and factor and it returns the finished result without charging again. Prices: megapotato.app/pricing; the API itself: megapotato.app/developers.
In the benchmark's price comparison (100 photos a month, three sizes averaged, prices as of 14 October 2026) the MegaPotato API came to $0.104 a photo, the cheapest of the four leaders: Topaz Wonder 3.5 was $0.24 through fal.ai and $0.54 through Topaz's own API, Crystal $0.35. The full comparison.
The photo goes to our server for the job, stays there 90 days so you can download it again, and stays yours (Terms, Privacy Policy).
