Rikan I2V Painter (Tiled VAE)

This is the sibling of the Rikan I2V Painter with one job's worth of difference: instead of encoding your start image with the stock vae.encode, it uses vae.encode_tiled - so the first-frame encoding splits into spatial and temporal tiles instead of hitting the whole clip's latent at once. Same conditioning outputs, same motion_amplitude magic, same Wan latent shape, just a VAE encode that won't blow up your VRAM when your clip is 720p and long.

Wan encoding at high resolution is a real memory wall. The community's standard workaround for decode OOMs is to swap in the tiled VAE decoder, and the same logic applies on the encode side: a big start frame or a long clip means a big intermediate latent, and on 8–12GB cards the stock encode can be the thing that dies before you ever reach the sampler. This node is that fix applied where the pack does its own encoding. If you're running the plain Painter happily, you don't need this. If you've seen an encode crash on 720p+, switch.

What the extra inputs do

The tiled variant keeps every input from the base Painter - positive, negative, vae, width, height, length, batch_size, motion_amplitude, color_protect, correct_strength, plus optional clip_vision and start_image - and adds four:

• tile_size (default 512) - how big each spatial tile is. Smaller = less memory, more seams to blend.
• overlap (default 64) - how much neighboring spatial tiles share, so the seams don't show.
• temporal_size (default 64) - the tile length in frames for the temporal dimension.
• temporal_overlap (default 24) - frame overlap between temporal tiles.

Outputs are identical to the base Painter: high_positive, high_negative, low_positive, low_negative, and latent. Wire the low_* pair into your KSampler for the motion-amplified anchor, or high_* if you want the first frame held dead-still.

The trade-off to respect

Tiling is not free. Smaller tiles and more overlap mean more encode passes and slower runs - you're trading VRAM for time. The defaults (512px tiles, 64px overlap, 64-frame temporal tiles) are a sane middle ground; if you're not OOMing, leave them alone, because fiddling with tile sizes to squeeze out a few extra frames of speed is a classic way to introduce visible tile boundaries in your first frame. And remember the rest of the pipeline: this only fixes the encode. When the KSampler finishes, the decode is a separate step - pair it with the pack's Wan Spatio-Temporal Tiled VAE Decode (or ComfyUI's built-in tiled decode) if that's where your card gives up.

One honest note: the tile parameters only govern the start-image encoding, not the sampling or decode. If a clip still OOMs, look at the sampler stage too - this node fixes exactly one step of the pipeline.

Install is the standard pack route: ComfyUI Manager → search Rikannodes, or clone into custom_nodes, install requirements.txt, restart. It's under Rikannodes in the menu. Same caveats as its sibling: Wan-specific latent shape, and motion_amplitude past ~1.5 starts looking pushed.