WanCameraImageToVideo (Tiled VAE encode)

Want the camera to dolly, pan, or zoom through your Wan render instead of the shot staying static? That's what ComfyUI's WanCameraImageToVideo node is for, and WanCameraImageToVideo (Tiled VAE encode) is its drop-in replacement with the VAE encode run in tiles - spatial tiles plus temporal frame-chunks, so the encode phase doesn't eat your VRAM budget before the denoiser even starts.

How it works

It's structurally the plain I2V node (start frame → concat_latent_image conditioning, plus optional CLIP Vision) with one extra wire: a camera_conditions input of type WAN_CAMERA_EMBEDDING. That embedding describes the camera trajectory and gets attached to both the positive and negative conditioning so the model knows how the viewpoint moves over the clip. In current ComfyUI, the source of that embedding is the core WanCameraEmbedding node, which takes a simple pose choice - Static, Pan Up/Down/Left/Right, Zoom In/Out, or rotation - along with width, height, length and speed, and hands you the embedding plus the matching dimensions. Wire its output into camera_conditions here.

Inputs that matter

• camera_conditions - the WAN_CAMERA_EMBEDDING from a camera-trajectory node. The whole point of this node.
• start_image - your first frame.
• clip_vision_output - optional CLIP Vision conditioning.
• The standard width / height / length / batch_size, and the four tiling knobs tile_size / overlap / temporal_size / temporal_overlap.

Outputs: positive, negative, latent - sampler conditioning plus samples.

Install

The pack installs as one unit:

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

or via ComfyUI Manager (search "WanImageToVideoTiled"), then restart. No extra dependencies - it only needs your existing Wan camera-capable checkpoint and its VAE.

The honest caveats

The README lists this one as unchecked - the author hasn't personally tested it, unlike the I2V, Fun Control, FirstLastFrame, VACE, and Wan22 nodes. It's also conditional on the camera-embedding node being present in your ComfyUI build; if you're on an older release where it doesn't exist, this node has no practical input to work with. Two pack-wide reminders: tiling only relieves the encode stage (the 27B camera-control denoiser is still a VRAM hog), and pair the front end with VAE Decode (Tiled) on the way out - decode is usually the slowest step of a Wan render.