WanAnimatePlus ClipVisionEncode V2

The original WanAnimatePlus ClipVisionEncode is built around a single reference image with an optional second slot. Fine when your identity anchor is one picture. Painful when it's a dozen. V2 fixes the obvious thing: instead of image_1/image_2, it takes one images input and encodes every frame in the IMAGE batch in a single pass. One node, one wire, as many references as you care to hand it.

What changes. Everything else is the same muscle memory. You still pass a clip_vision model, and the output is still WANVIDIMAGE_CLIPEMBEDS for the clip_embeds input on WanAnimatePlus AnimateEmbeds or the SCAIL-2 embeds node. The strength slider is now a single multiplier applied to every embed in the batch - no per-slot strengths, which is the one thing you give up versus the original. crop (center 224x224 vs disabled), combine_embeds (average/sum/concat/batch), negative_image (now accepts an image batch for the uncond side), and tiles/ratio for matteo's tiled encoding all carry over unchanged.

That combine_embeds choice matters more here because you actually have a batch to combine. average blends all your reference frames into one embed - the classic multi-angle identity anchor. concat stacks them, which keeps their distinct features but makes the embed bigger downstream. batch keeps every embed separate, which is what you want when the downstream node expects per-frame conditioning rather than a merged identity. If your face still drifts between shots, feed it five frames of the character from different angles and let average glue them together.

When to reach for it. Honestly, whenever you'd otherwise chain several of the single-image node. Multi-shot sequences, an image sequence extracted from a reference clip, or SCAIL-2 jobs where you're anchoring multiple subjects - one batch in, one clean embed out, and your graph stays readable instead of sprouting a row of encoder nodes. The tooltip says it plainly: all frames in the IMAGE batch are accepted, no per-frame workarounds.

The same caveats as its sibling. Use the CLIP vision file that matches your checkpoint - mismatched encoders silently change identity behavior. Don't lean on strength to fix drift; it makes output mushy past ~1. And tiles is a slower-but-sharper option, so leave it at 0 unless your reference resolution actually warrants it.

Install the pack through ComfyUI Manager (search "WanAnimatePlus") or git clone https://github.com/wuwukaka/ComfyUI-WanAnimatePlus into ComfyUI/custom_nodes, restart, and remember the fork-wide rule: build the whole chain from WanAnimatePlus nodes rather than mixing in the original WanVideoWrapper ones. It's a small upgrade, but if you've ever stared at a wall of encoders, you'll appreciate it.