BerniniR · Load VAE (Wan)

BerniniR · Load VAE (Wan) loads the AutoencoderKLWan VAE out of the model bundle you got from BerniniR · Load Model. It's a small node with two jobs, and both matter: it feeds BerniniR · VAE Decode so your latent becomes frames, and it feeds BerniniR · Encode Source/Reference so your source video and reference images become the latents the sampler can edit against.

On this (diffusers) path there's no separate VAE download - the bundle the Load Model node fetched already contains it, and the model_path output tells this node exactly where to look. That's the whole point of the BR_PATH wiring: one source of truth for all weights.

Inputs

• model_path (BR_PATH) - the model_path output from BerniniR · Load Model. Required.
• dtype - fp16 (default), bf16, or fp32. fp16 is the default for good reason: the Wan VAE runs fine in fp16 and saves VRAM. If you ever see artifacts that look like precision loss in decoded frames, bf16/fp32 is the sanity check - but don't start there.

Output

vae (BR_VAE) - this pack's VAE type. It's not a stock ComfyUI VAE, so feed it to this pack's Decode and Encode nodes, not to the core VAEDecode. (On the GGUF/native path, you'd load the Wan VAE with ComfyUI's native VAE loader instead - this node is for the dual-expert diffusers path.)

One thing to know about the Wan VAE

The Wan VAE compresses time ~4×: N frames become ((N-1)//4)+1 latents and decode back to (T_lat-1)*4+1 frames. That's why the community's "4n+1" arithmetic keeps coming up - 111 frames in → 109 out unless you pad. The pack's own defaults (81 frames) are already 4n+1-safe, so don't fight it on the happy path. It only bites when you chain or extend beyond the native window.

Install & troubleshooting

Same as every node in the pack: ComfyUI Manager, or clone into custom_nodes and pip install -r requirements.txt. If you see a BR_VAE socket that won't accept a connection, you've grabbed the stock VAE from the wrong loader - it has to come from this pack or the matching native loader. And if the node errors with a missing-weights message, it's because the Load Model node hasn't downloaded the bundle yet (first run is ~40 GB); run it once and let auto_download finish.