IPAdapter Load Embeds

This is the other half of Save Embeds. Where Save writes an encoded reference to disk, Load Embeds reads it back - you pick a saved embedding from a dropdown and it hands you an EMBEDS object to feed into an apply node. Together they let you encode a reference once and reuse it across as many generations as you like, without re-running CLIP vision every time.

It's a small node with a clear job. You reach for it when you've got a cached reference "look" or "face" you want to drop into a new workflow, or when you're sharing embeddings between graphs. Because the embedding is precomputed, applying it is quick and perfectly repeatable - the same file always produces the same conditioning.

How it works

There's genuinely nothing to configure. The node scans ComfyUI's saved-embeds location, lists what it finds in a dropdown, and outputs the one you select. All the real work - the encode - already happened when the embed was created. This node just deserializes the file and puts an EMBEDS on the wire.

The input and output

• embeds - a dropdown of the saved embedding files it found. If the list is empty, that's not a bug - it means there are no saved embeds yet. You have to create one first (encode a reference and run it through IPAdapter Save Embeds), then this dropdown will populate.
• Output EMBEDS - the loaded embedding. Wire it into the pos_embed (or neg_embed) input of an apply node like IPAdapter Embeds Batch, and you're applying a cached reference.

That's the whole node: pick a file, get an embedding.

Installing the pack

ComfyUI Manager: search ComfyUI IPAdapter plus, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/cubiq/ComfyUI_IPAdapter_plus

then restart, and keep ComfyUI current. Loading an embed doesn't need model files itself, but the apply node you feed it into needs the adapter weights in ComfyUI/models/ipadapter, and creating embeds in the first place needed the CLIP vision encoders in ComfyUI/models/clip_vision.

Where people get burned

The number-one "why is the dropdown empty" moment: there's nothing saved yet. Load Embeds can only show files that already exist. Run a reference through an encoder and IPAdapter Save Embeds first, and the file appears here. If you saved one and it still doesn't show, the embed landed somewhere the node isn't scanning - check that Save Embeds wrote to ComfyUI's expected output area, and refresh the node.

The second is the family-mismatch trap that follows embeds around: an embedding encoded for one setup (SD 1.5 ViT-H, say) will throw a shape error when applied against a different-family adapter (SDXL bigG). Loading is fine - the mismatch surfaces at the apply node - so if you get a tensor-size error downstream from a loaded embed, that's the cause. Keep your saved embeds labeled by family. And the standing note: this pack targets SD 1.5 and SDXL only, in maintenance mode.