Anima Image Embedding Loader (Legacy)

ComfyUI pack authors love leaving the old version around with a "(Legacy)" suffix and this is exactly that. Anima Image Embedding Loader loads a pre-computed image embedding from a .pt file and hands you an ANIMA_IMAGE_EMB. It was built for the pack's earlier, Qwen3-VL-based pipeline. The shipped code has since moved to SigLIP2, the Apply node now takes SIGLIP_FEATURES, and nothing in the current pack consumes this node's output type.

So the straight answer: you almost certainly don't need this node. If you're setting up Anima IP-Adapter fresh, build the Load Image → Anima SigLIP2 Encode Image → Anima IP-Adapter Apply chain and don't look back.

When you'd actually touch it

Three situations, and they're all edge cases:

1. You have a stack of pre-computed embeddings from the old pipeline (saved as 1024-dim L2-normalized Qwen3-VL vectors via the README's precompute script) and want to reuse them without re-encoding. Fine - this loads them, squeezing to 1D if they came in batched.
2. You wrote or use a custom node that consumes the ANIMA_IMAGE_EMB type and want to feed it from a file.
3. You're on an older snapshot of the pack where the Apply node still had an image_emb input instead of siglip_features. On today's build that socket is gone, so this loader's output has nowhere to go.

The mechanics

One required input, emb_path, a plain string path to the .pt file. It loads with torch.load(..., weights_only=True) on CPU and squeezes extra dimensions. Output is ANIMA_IMAGE_EMB. If you're wondering what format the .pt needs to be - it's whatever the old precompute script saved: a single tensor, typically [1, 1024], from pooling a Qwen3-VL embedding and L2-normalizing the first 1024 dims.

That last part matters if you're tempted to feed it SigLIP2 features: don't. SigLIP2 patch features are a different shape, a different encoder, and a different type. This node is for Qwen3-VL-era .pt files only, and mixing them up is how you end up with the Apply node refusing the connection and a head-scratching afternoon.

Install and the usual gotcha

Installs with the pack, nothing extra:

cd ComfyUI/custom_nodes
git clone https://github.com/Wenaka2004/comfyui-anima-ipadapter
pip install -r comfyui-anima-ipadapter/requirements.txt

The README hasn't caught up with the code - it still documents the Qwen3-VL flow this node belongs to, while the actual nodes now speak SigLIP2. So if the README's Option A ("pre-computed embedding, no extra VRAM") reads like it should Just Work and then doesn't, that's the rot. Trust the node graph, not the README. And if a workflow you downloaded uses this loader feeding into the current Apply node, the author of that workflow is on the old build - update your pack and swap in the SigLIP2 encode path.