WanAnimatePlus VAELoader

Every video starts as pixels, gets crushed into latents by the VAE, and comes back out through it at the end. In the WanAnimatePlus universe, this is the node that loads that compressor - a Wan VAE from ComfyUI/models/vae - with a couple of options that matter more than you'd expect for a "just load it" node.

The reason this fork bothers to re-export a VAE loader under its own prefix is the same reason it re-exports everything: the pack's chains use WANVAE as a type, and mixing in an official VAE loader means type mismatches or subtle object differences. Use the pack's loader with the pack's nodes and the cables line up. The mechanism is the standard one - load the checkpoint, handle the model. state-dict prefix, and hand back a WANVAE.

The knobs that matter

• model_name - the VAE checkpoint, from the ComfyUI/models/vae folder. For Wan this is the 3D causal VAE, and it's also what SCAIL-2 and the Wan Animate chains share.
• precision - fp16, fp32, or bf16 (default). The VAE is usually fine at bf16; if you see color or banding artifacts, step up to fp32 rather than chasing it in the sampler.
• compile_args - optional; wire the output of WanAnimatePlus TorchCompileSettings here to compile the VAE too. Worth it if you're doing long encodes/decodes repeatedly.
• use_cpu_cache - off by default, and the tooltip is honest: it "reduces VRAM usage, but slows the VAE down a lot." The classic trade. Flip it when the VAE is what's blowing your VRAM budget.
• verbose - on, logs memory usage while the VAE runs. Handy for a one-time look at what's eating your card, then turn it off.

Output is a single vae (WANVAE) that feeds every node in the pack that encodes or decodes: WanAnimatePlus AnimateEmbeds, SCAIL_2 Embeds, Bernini, VAE Decode, the Flow chain, all of them.

Wiring it

In a typical WanAnimatePlus SCAIL-2 workflow it feeds both the embeds node (which VAE-encodes your reference image, pose, and masks) and the decode end. If you're on the official-compatible Flow chain, note that WanAnimatePlus SCAIL_2 Flow Embeds takes a standard VAE socket - the pack's own WanAnimatePlus VAELoader works there because the Flow nodes are the official-typed bridge. And if you're decoding long clips, the tiled_vae option lives on the decode node, not here.

Install

Part of ComfyUI-WanAnimatePlus, so:

cd ComfyUI/custom_nodes
git clone https://github.com/wuwukaka/ComfyUI-WanAnimatePlus.git

Restart ComfyUI, or install via Manager by searching "ComfyUI-WanAnimatePlus". Keep the original ComfyUI-WanVideoWrapper installed per the README, drop your Wan VAE into ComfyUI/models/vae, and stick to the pack rule: one node family per workflow.

The honest summary: it's a loader, not a party. But it's the loader that keeps the pack's type system coherent, and the use_cpu_cache and precision dials give you the two levers you'll actually pull when the VAE becomes your memory bottleneck.