Magnific Precision Upscale

If Creative is the Magnific node that redraws your image with better detail, Precision is the one that's supposed to leave your image alone. Same service, same key, opposite deal: you're buying cleaner edges and a bit of gentle enhancement, not a different picture.

That distinction sounds obvious and then people mix the two up, get a portrait back with someone else's face, and decide the whole service is a hallucination machine. Magnific's own two engines are honest about the difference; the slider names tell you which one you're in.

What you're choosing between

Two model entries, and the parameters change with the pick:

Precision V2 - takes a scale_factor (2x/4x/8x/16x), a flavor (sublime, photo, or photo_denoiser - read those as processing styles; denoiser is the one for a source that's noisy or has been through a bad JPEG), plus the three detail sliders: sharpen (edge sharpness, 0–100, defaults low at 7), smart_grain (grain to stop the result looking like it was ironed, default 7) and ultra_detail (fine added detail, default 30).

Precision V1 - the older engine. Same three sliders and no flavor or scale factor; sharpen defaults to 50 rather than 7, which is the main behavioural difference you'll feel. If you port a V1 setting to V2 and the result looks soft, that's why.

Common to the node:

• image - the image to process. A single IMAGE input, not an autogrow socket.
• model - the dynamic dropdown holding the entries above, plus Custom model for an ID.
• filter_nsfw - boolean, default off, advanced.
• extra - JSON merged into the request.
• retry_count - total attempts, ~3s apart.
• provider - optional; defaults to Magnific, keyed from api_key.json's magnific entry or MAGNIFIC_API_KEY.

Output: one image.

The honest comparison

This is where a bit of ecosystem context saves you money. Upscaling splits into three jobs, and Precision is the "more pixels, more cleanly" one - which is exactly the job where local tools have become genuinely good and free. ESRGAN-family models cost you a few seconds and nothing else; NVIDIA's hardware path is 4GB of VRAM and seconds per image; the modern open detail upscalers like SeedVR2 do detail work locally at a quality the community rates above the paid tier for restoration. The KB's upscaling doc has the whole ladder, and the summary is unsentimental: an author shipping an open alternative to Topaz reported dropping their subscription over it, which is where the direction of travel sits.

So what's Precision for? Sources the local models mangle - awkward textures, mixed content, a photo with real grain structure that needs preserving rather than smoothing - and jobs where you want a well-tuned grain and denoise pair applied by a service that's had years of iteration on exactly that. If your images are AI renders at 1024 and you just want 4K, you probably don't need this node. Try the local path first; you can always pay later.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtNodes
cd ComfyUI-YogurtNodes
pip install -r requirements.txt

ComfyUI Manager installs it too - search "YogurtNodes"; the node appears as "Magnific Precision Upscale (Yogurt Nodes)". Key file:

{ "magnific": "your-api-key" }

at custom_nodes/ComfyUI-YogurtNodes/yogurt_nodes/llm/api_key.json. The Provider: Magnific node adds a proxy field and a task timeout that defaults to 900 seconds - worth wiring if you're doing 8x or 16x jobs, because those are the runs that outlive a short wait.

Bruises

Sharpening a source that's already sharp. You get halos and a crunchy texture. Precision at high sharpen on a good render is a downgrade; that's what the low default on V2 is telling you.

Running out of patience. Precision is not instant the way a local ESRGAN pass is. The job polls, and at 16x on a big source you're in minutes territory. Do not set retry_count above 1 on a slow job you're still learning - you'll pay for the same timeout three times.

Expecting Creative's look. If your image came back unimproved rather than changed, you bought the right product. Precision's whole value proposition is that it doesn't invent.