Fal Upscale

Local upscaling is solved - if you have the VRAM and the right model wired up. Fal Upscale is the version for when you don't want to be the upscaler-adjacent-system administrator of your own machine: drop the node, wire in your image, pick Real-ESRGAN or one of the hosted upscalers, and get a bigger, cleaner image back as a normal IMAGE tensor.

It's one of eleven nodes in ComfyUI-Fal-Gateway, and it's the pack's answer to the "more pixels" and "more detail" jobs at once. The dropdown auto-detects the upscale-shaped endpoints in fal's catalog - Real-ESRGAN (the ESRGAN-family classic), Clarity Upscaler, Crystal Upscaler - and exposes whichever are live.

Inputs and outputs

• image - the thing to upscale, from LoadImage.
• model_id - dropdown of the upscale endpoints.
• prompt - required widget that most plain upscalers ignore. Leave it empty for Real-ESRGAN-style pixel enlargement; the node only sends a prompt you actually type.

Outputs: image (the upscaled IMAGE tensor → SaveImage), image_url (STRING), info (STRING).

How it works

Your image uploads to fal, the upscale endpoint runs, and the result downloads and decodes back into a tensor. The mechanism is deliberately boring - which is the point. The job it's solving is a hassle-removal, not a capability unlock: upscaling models are open and run fine locally, but this moves the work off your card entirely. Worth keeping the KB's distinction in your head before you pick: ESRGAN-family models add pixels (and can't invent detail), while Clarity and Crystal are the generative side that rebuilds and adds texture. If you want the generative restoration, pick accordingly - a naive "bigger = better" grab at the cheapest model is the classic mistake.

Install

With the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/modbender/ComfyUI-Fal-Gateway
pip install -r requirements.txt
export FAL_KEY=your_fal_key

Restart ComfyUI; it's under Fal-Gateway. No model downloads - the upscale models live on fal's side.

Gotchas

• You're paying for something you could do locally. If you have a working ESRGAN or SeedVR2 pipeline, this node is strictly a convenience (and a per-call cost). Its real audience is the no-GPU or low-VRAM workflow, or the pipeline that already reaches into fal for its generation and wants the upscale in the same graph, same bill.
• Fresh nodes can show <no models available> in the dropdown until the catalog fetch finishes; right-click → "Fal-Gateway: refresh catalog cache" if it's stuck.
• No key → "FAL_KEY not set" hard error, fix spelled out in the message.
• Watch the cost badge - generative upscalers bill more than the pixel-enlargement ones, and the badge in the title bar reflects that before you queue.

Honest verdict: it's the least glamorous node in the pack and the most honest use of it - a one-node upscale step that slots into a fal-backed workflow without touching your VRAM. If you're already paying for fal generation, folding the upscale in costs less hassle than installing another local chain.