WanVideo SVIPro Embeds

Wan's native context is 81 frames. That's roughly five seconds, and it's the wall every local video person hits. SVI - Stable Video Infinity, EPFL's error-recycling LoRA - is the trick that blew past it, producing 20-second and even minute-plus continuous takes on plain Wan 2.2. The January 2026 flagship thread that put it on the map showed a 1280x720, 20-second continuous clip generated in 340 seconds, and the community reaction was not subtle. This node is the piece that makes the extension actually chain: it builds the WANVIDIMAGE_EMBEDS that carry each chunk's anchor forward into the next, so the SVI 2.0 Pro LoRA has something to condition on.

If you're wiring up an SVI Pro workflow, this is the embeds node in the middle of the loop.

How it works

SVI works by teaching Wan to survive its own mistakes - autoregressive video breaks because the model trains on clean frames but then has to condition on its own error-laden output, and SVI's fine-tuning feeds those errors back as supervision. But the plumbing still has to hand the model the right anchor for each window. That's this node. You give it an anchor latent, tell it how many frames the chunk is, and - for continuation - the previous chunk's samples, and it assembles the image embeds the sampler needs to keep the shot going instead of starting fresh.

The inputs and outputs that matter

• anchor_samples (LATENT, required) - the latent this chunk is anchored to. This is what keeps the extended segment tied to your source instead of drifting into a new scene.
• num_frames (default 81, step 4) - the frame count for this window. It steps by 4 because the Wan VAE compresses time roughly 4x, so frame counts want to land on the 4n+1 grid.
• prev_samples (LATENT, optional) - the previous chunk's latent. Feed this on every window after the first; it's the handoff that makes the clip continuous rather than a cut.
• motion_latent_count (default 1) - how many motion latents carry over from the previous segment. More overlap generally means smoother joins.

Output is image_embeds (WANVIDIMAGE_EMBEDS), which goes into WanVideoSamplerv2.

How to install it

ComfyUI Manager - search ComfyUI-WanVideoWrapper, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper
pip install -r ComfyUI-WanVideoWrapper/requirements.txt

then restart. This node needs the SVI 2.0 Pro adapters - the HIGH/LOW rank-128 LoRA pair Kijai hosts on Kijai/WanVideo_comfy - loaded onto a Wan 2.2 I2V-A14B base. SVI is a LoRA, not a model, so you're adding two ~1.1GB files to an existing Wan workflow, not swapping the model.

Common issues & troubleshooting

The clip goes progressively blurry the longer it runs. This is SVI's best-documented failure mode, and it's worse when you stack motion LoRAs on top. It compounds with length. There's no clean fix - it's the price of the technique - so keep segments to the reliable 20-25 second band rather than chasing the extreme edge.

Faces change from shot to shot. SVI conditions length, not identity. The top critical reply on the flagship thread was literally "why are her facial expressions changing every half second?" If you need one person to stay one person, add a character LoRA to the base model; this node won't hold identity for you.

It crashes past a certain length. People routinely find the ceiling is ComfyUI itself dying during stitching and assembly, not the adapter or VRAM. One user did 1.5 minutes perfectly and anything longer crashed Comfy. If you're hitting that, stitch the chunks outside ComfyUI instead of asking it to assemble a giant clip in one go.