Magnific Image Upscale (Creative)

This is the node version of the product that SUPIR was famously positioned against. Magnific Image Upscale (Creative) is Magnific's prompt-guided upscaler: it doesn't just interpolate pixels, it generates new detail - texture, lighting, sometimes whole new micro-features - while enlarging your image 2x, 4x, 8x, or 16x, up to a 25.3-megapixel output ceiling. In ComfyUI it's a partner node: cloud-hosted through Comfy's proxy, billed per call against your Comfy account, zero install.

The "Creative" is doing real work here. Where the Precise V2 sibling is about sharpness and grain control, this node is where you tell the upscaler what kind of detail to invent. Write a prompt describing the material or mood - "oil painting texture," "sharp editorial photography" - and the upscaler leans into it. If you've read the KB's upscaling essay, this is the generative-restoration end of the spectrum, the one that adds detail rather than preserving it.

How it works

Your image is uploaded to Magnific's image-upscaler endpoint; the node polls until the job finishes and downloads the result as a normal IMAGE tensor. The four sliders are the creative core:

• creativity (−10 to 10) - how much new stuff the model is allowed to invent.
• hdr (−10 to 10) - level of definition and detail, per the tooltip.
• resemblance (−10 to 10) - how closely it must stay faithful to the original.
• fractality (−10 to 10) - prompt strength and intricacy per square pixel. This one's the flavor knob; push it for more of whatever your prompt describes.

Plus optimized_for - a preset list (soft_portraits, art_n_illustration, films_n_photography, videogame_assets, etc.) that pre-tunes the whole thing for the subject, and an engine selection (automatic, or Magnific's named engines like illusio, sharpy, sparkle) tucked in as advanced options.

The inputs that matter

• image - single image only; batches error out.
• prompt - the detail you want invented. Empty is fine; filled is where the magic is.
• scale_factor - 2x through 16x. Bigger costs more, and 16x needs a small-ish input to stay under the pixel cap.
• creativity / resemblance / fractality / hdr - the four sliders. Defaults are all 0; move them in the direction of the problem (stiff result → up creativity, drift from original → up resemblance).
• auto_downscale - off by default. If your input is large and the requested scale would blow past 25.3MP, the node errors unless this is on, in which case it shrinks the input first to fit.

Output is IMAGE, ready to save.

Where people get burned

• The megapixel cap, every time. It's in the node description, and still, the first big 16x attempt fails for everyone. Flip auto_downscale on when going big.
• Single image only. Feed it a batch and it stops you cold.
• "Creative" means it changes things. This isn't a restoration tool. If you need pixel-faithful enlargement, the Precise V2 node (or a local ESRGAN-type upscaler) is the right tool. Creative is for when you want the upscaler to reinterpret.
• Cost. Magnific sits on the expensive end of the partner family, and the price scales with output size - the badge shows a range, and 16x of a large image is real money. Iterate on a 2x pass before committing to the 16x final.