Magnific Library Reference

If you've trained a character, a style, an element or a location in Magnific, this is how it gets into a generation you're running from ComfyUI. It's the node equivalent of typing @name in Magnific's web app - which is exactly what it produces.

It's a small node that does one job: pick an asset, output a reference bundle, wire the bundle into Magnific Generate Image or Magnific Generate Video. The reason it deserves its own article is that slot-based consistency is the thing hosted services still do better than most local setups, and this is the adapter that lets your local masking and compositing live downstream of it.

What you set

catalog is My Library (your own assets and anything shared with you) or Magnific (the public catalogue). type narrows the list - All, Characters, Styles, Elements, Locations.

asset is the picker, and it renders as @name (Type · Origin) grouped by type. Only assets in a ready state appear; something still training simply won't be in the list, which is the intended behaviour rather than a bug. If you created it a minute ago, press Refresh library on the node (or R) and it'll show up.

chain is the trick for attaching several assets to one generation: wire one Library Reference node's output into the next node's chain input, and the last one into the generation node. Then remember the ceiling - a generation takes at most 12 references, Library and reference images combined, and the nodes check before uploading anything rather than failing halfway.

The single output is library_references, which only the Magnific generation nodes understand - it's a MAGNIFIC_LIBRARY_REFS bundle, not something you can inspect.

The detail that matters most

Write the @name token in the prompt yourself, where you want the asset to appear. The tooltip is explicit that if you forget, the node appends the missing tokens to the end of the prompt - the reference still rides along, but its position in the sentence is gone. For a character that's usually fine. For "a @style render of a person" versus the token dumped at the end, it changes what the model conditions on.

Second, not every model takes every type. Video models accept characters, elements and styles but never locations. Image models list what they accept. The node stops with a message naming the model and the type before anything is sent, so the fix is to pick a different model or set the model to auto and let Magnific choose.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git

Or install "CustomNodePacks" from ComfyUI Manager, then restart. The Magnific nodes come with the pack - they're a port of Magnific's vendor pack magnific-comfyui 0.7.0, including the front-end bits that draw the asset list. Sign in first (Magnific menu, or the Magnific Save To node's sign-in button); without it the asset dropdown shows the sign-in placeholder and nothing else. Credentials go to ~/.magnific/comfyui_auth.json.

If the dropdown is stale after you train something in the browser, R is the answer, not a ComfyUI restart.

When this is worth it

Local consistency tooling has come a long way - IP-Adapter, LoRAs, face-swap nodes - and for a single character you've already got a LoRA for, dragging in a cloud generation just to use a Library asset is hard to justify. Where it earns its place is the case the local stack still handles badly: a studio of several trained elements and locations that need to co-occur in one shot, across models you don't have weights for. That's a slot-filling problem, not a prompting problem, and hosted services are good at it.