Freepik Upscaler Creative

Freepik Upscaler Creative is the node that sells this pack. It's a hosted, prompt-guided upscaler that does 2x through 16x, and the author's pitch is that it runs on the same Magnific technology people currently pay $39 a month for - except here you pay per upscale and nothing when you're idle. For low volumes, that math works out dramatically cheaper, which is exactly the framing the README leans on.

Before you buy in, get the job straight. This is the "invent detail" rung of upscaling, not the "add pixels" rung. The community spends a lot of words on this distinction because two people doing different jobs keep comparing results: an upscaler that adds pixels (ESRGAN, Lanczos) never hallucinates; a generative one like this exists to invent plausible detail that was never in the source - rust on metal, brick texture, fabric weave. Creative is firmly the second kind. That's its strength, and it's also the trap.

How it works

The node base64-encodes your image and posts it to api.freepik.com/v1/ai/image-upscaler, then polls the async task until it's done - up to a ten-minute ceiling. It has two genuinely thoughtful features for a paid node. First, flip confirm_before_run on and a dialog pops up on the canvas with the estimated cost before anything billable happens - worth doing until you trust your settings. Second, the pack's web extension paints a live status bar on the node, so you can watch it poll instead of staring at a frozen graph. Results are cached by parameters + seed + image hash for 30 days, so identical reruns are free.

The inputs that actually matter

The parameter list is long, but the ones you'll actually touch:

• upscale_factor - 2x, 4x, 8x, 16x. The price climbs fast with this.
• prompt - optional, but this is your lever for what detail gets invented. Feed it "enhance architectural details, photorealistic render, natural lighting" and it steers the enhancement.
• creativity - integer slider, −10 to 10. This is the big one. Start around 3–5 for a believable amount of added detail and dial from there.
• hdr, resemblance, fractality - also −10 to 10. Resemblance pulls the result back toward the original; crank it if the output starts drifting from your source.
• optimized_for - nine presets (soft/hard portraits, 3D renders, videogame assets, sci-fi/horror…). Picking the right one matters more than fine-tuning the sliders.
• engine - automatic plus three named Magnific engines (illusio, sharpy, sparkle). automatic is a fine default.
• seed - reproduce a look by keeping it fixed; note the cache key includes it.

One warning: the README's parameter table claims creativity runs 0.0–1.0. The node itself exposes −10 to 10 integers, matching Magnific's own scale. Trust the node, not the README.

Outputs are upscaled_image, info (a status string), cost (€ float), and seed (so you can freeze what worked). Cost tiers by output area run from about €0.10 at 2x up to €1.20–1.60 for enormous 16x outputs - that cost output exists for a reason.

Installing it

ComfyUI Manager → search ComfyUI-Freepik, or:

cd ComfyUI/custom_nodes
git clone https://github.com/febogallo/ComfyUI-Freepik
cd ComfyUI-Freepik
pip install -r requirements.txt

Restart, grab an API key from the Freepik Developer Dashboard, and it's under Freepik/Editing. No model downloads - this one is all server-side.

Where people get burned

Remember what rung you're on. Creative will rewrite faces and invent detail that isn't there - that's the design, not a bug, and it's why the local community's generative alternatives (SUPIR, SeedVR2) come with the same warning. If you need fidelity over wow, that's what its sibling, FreepikUpscalerPrecision, is for. And mind the money: 8x/16x reruns with a fresh seed are the fastest way to watch €1.60 calls stack up. Keep confirm_before_run on, keep your seed stable when you're iterating, and let the cache do its job.