Magnific Image Style Transfer

Style transfer, done as a hosted API call: Magnific Image Style Transfer takes two images - the one you want restyled and a reference whose look you want to steal - and produces a hybrid, with the degree of "keep my original structure" vs "copy their style" exposed as separate sliders. It's a ComfyUI core partner node, no install, cloud-hosted through Comfy's proxy, billed per call against your Comfy account.

The reason this is worth reaching for over local style-transfer hacks: the two-strength control is actually meaningful. style_strength says how much of the reference's look to apply; structure_strength says how much of the original image's composition to keep. You can push the style hard while holding the structure at 50%, which is the sweet spot for "same photo, different aesthetic" - exactly the thing that's fiddly to tune with a local model and a denoise slider.

How it works

Both images get uploaded to Magnific's style-transfer endpoint; the node polls the job and returns the result as a regular IMAGE tensor. The fun is in the presets: flavor picks the vibe (from faithful through gen_z and psychedelia to detaily), and engine picks the rendering engine (including a 3d_cartoon option, which is as fun as it sounds). There's also an optional portrait mode that unlocks face-aware treatment - facial beautification and portrait-specific styling.

The inputs that matter

• image - the photo you're restyling.
• reference_image - where the style comes from. Both must be single images; batches error.
• prompt - optional guidance layered on top.
• style_strength (0–100, default 100) - how much of the reference's look.
• structure_strength (0–100, default 50) - how much of the original's composition survives.
• flavor - the style preset: faithful, gen_z, psychedelia, detaily, clear, or donotstyle variants.
• engine - balanced, definio, illusio, 3d_cartoon, colorful_anime, caricature, real, super_real, softy.
• portrait_mode - dynamic combo; enabling it adds portrait_style and portrait_beautifier options for face work.
• fixed_generation - default on. Flip it off and every run gets a degree of randomness, which is great for exploring variations and bad for reproducibility.

Output is IMAGE, ready to save or feed downstream.

Where people get burned

• Single-image inputs only. A batch into image or reference_image throws immediately. Check your batch isn't 4 frames before queueing.
• The structure/style tension. Both at 100 means the model is being told to do two contradictory things; the defaults (100 style / 50 structure) are a sane place to start and usually where you should end up.
• portrait_mode hides options. Enable it and new dropdowns appear; that's a feature, not missing inputs.
• Cost. Style transfer here is a flat per-call charge that sits above the cheap end of the partner family. When you're exploring looks, disable fixed_generation and generate a couple of variants in one session rather than one at a time.