WanFunControlToVideo (Tiled VAE encode)

If you've seen the "Wan Fun Control" models - the per-frame control checkpoints that let you drive a Wan render with an entire video rather than just a first frame - this is the tiled-encode version of ComfyUI's core WanFunControlToVideo node. WanFunControlToVideo (Tiled VAE encode) is a drop-in replacement that runs the VAE encode in spatial and temporal tiles, so the fun-control pipeline doesn't blow your VRAM budget during the encode pass.

What it does differently

The core node is special because it encodes two streams into a 32-channel concatenated latent: your start_image in one set of channels and your control_video in the other, both run through the Wan 2.1 latent format (process_out). The tiled version does exactly that, except each stream goes through vae.encode_tiled() - spatial tiles plus frame-chunks - instead of one giant encode. Same outputs, same conditioning math, lower peak memory.

Inputs that matter

• start_image - your first frame. It gets upscaled to width/height and encoded into the concat_latent_image that anchors the start of the clip.
• control_video - the per-frame control signal. Feed it whatever structure you want the model to follow (pose, depth, edge, or any video you're using as a guide). This is what separates Fun Control from plain I2V.
• clip_vision_output - optional CLIP Vision conditioning, same as the other Wan nodes.
• width / height / length / batch_size - defaults 832×480, 81 frames. And the four tiling knobs: tile_size / overlap / temporal_size / temporal_overlap.

Outputs are the standard positive, negative, latent triple - wire positive and negative into your sampler's conditioning, and latent into samples. Nothing exotic to learn if you've run any other Wan conditioning node.

Install

The pack installs as one unit, so it's the same routine as every node here. ComfyUI Manager → search "WanImageToVideoTiled", or:

cd ComfyUI/custom_nodes
git clone https://github.com/stduhpf/ComfyUI--WanImageToVideoTiled

Restart, and you're done. No requirements.txt, no model downloads - it builds the concatenated latent with ComfyUI's own Wan 2.1 latent-format helpers, so it only needs your existing Wan Fun Control checkpoint and its VAE.

The honest caveats

The author lists this node as successfully tested, so it's a safe pick. The two pack-wide warnings still apply: tiling only relieves the encode stage, not the denoiser - Fun Control checkpoints are still 14B-scale models that eat VRAM on their own - and you should pair it with VAE Decode (Tiled) at the end of the pipeline, since decode tends to be the slowest step. Quality-wise, tiling is theoretically a hair below a full encode because of seam blending; in practice the author reports no visible difference. If you do spot tile lines, raise overlap before touching anything else.