IPAdapter Load Embeds V2

IPAdapter Load Embeds V2 is the reload half of the saved-embedding workflow. IPAdapter Save Embeds V2 writes an embedding to disk as a .ipadpt file; this node reads it back into the graph as an EMBEDS tensor, ready to feed IPAdapter Embeds V2. It's the node that turns "encode once, use forever" from a single-session convenience into a persistent asset library.

What it does

A single required input, embeds - a dropdown listing every .ipadpt file it finds under your ComfyUI input directory. The node walks that directory (including subfolders) and lists anything ending in .ipadpt. Pick one and you get a single EMBEDS output. That's the whole interface, and it's deliberately bare.

When you'd use it

Any time a reference embedding is worth keeping. The workflow: encode a reference with IPAdapter Encoder V2, save it with IPAdapter Save Embeds V2 (which drops files into your output folder by default), then move the .ipadpt somewhere under input and load it whenever you need that reference again. Because the file is a fixed tensor, every load is identical - no re-encoding variance, no crop drift between runs. If you generate from a brand style or a character reference repeatedly, this is how you stop re-encoding it each time. Load it, combine it with other embeds via IPAdapter Combine Embeds V2, or feed it straight into IPAdapter Embeds V2 and generate.

The file-placement gotcha

The node lists files from the input directory, but the Save node writes to the output directory. So the two nodes don't naturally see each other's files - you save, then move (or copy) the .ipadpt from ComfyUI/output into ComfyUI/input (or a subfolder) before Load can list it. That trips people up constantly, because the natural assumption is that Save and Load share a directory. Once the file is in input, a refresh of the dropdown (or a ComfyUI restart) picks it up.

Installing it

ComfyUI Manager: search "ComfyUI_IPAdapter_plus_V2". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/chflame163/ComfyUI_IPAdapter_plus_V2

Restart ComfyUI. No models needed for this node.

Gotchas

Beyond the directory mismatch, the main thing is compatibility: a saved embedding is only meaningful to the apply nodes if the CLIP vision model you load on the apply side matches the one that produced it. Keep encode/save/load/apply on the same vision model. And this pack is a renamed snapshot of cubiq's V2 codebase, so its behavior is pinned to that codebase - no upstream fixes without a manual sync.