WanPhantomSubjectToVideo (Tiled VAE encode)

WanPhantomSubjectToVideo (Tiled VAE encode) is the tiled-encode version of ComfyUI's core WanPhantomSubjectToVideo, the node for Wan's reference-subject ("Phantom" / "Subject") models. Where plain I2V gives the model one first frame to work from, this gives it a whole set of reference images of a subject - front, back, a pose, a costume - so the render stays consistent with a character that isn't just a single snapshot. If that's your workflow, the VAE-tiled version exists for the usual reason: the encode of those reference images runs in tiles instead of one big pass, so it costs less VRAM.

How it works

Every image you pass in gets upscaled to your width/height and encoded with the tiled VAE, then the encoded frames are concatenated along the time axis into a single concat_latent_image. That's injected into the conditioning via time_dim_concat - so the model sees the subject's appearance as context that runs through the whole clip, not just as a first frame. The images input is a list, so you can hand it multiple references and it stitches them all into the timeline.

The output list is where this node gets interesting. You get positive, then two negative outputs: negative_text and negative_img_text. Looking at the source, negative_text is your original text-negative but carrying the real reference images, while negative_img_text is the negative with the image-concat zeroed out. In plain terms: one negative leans on the text, the other leans on the images - try both and see which stabilizes your subject without flattening the motion. Wire one into your sampler's negative conditioning and see.

Plus the usual positive/text inputs, vae, width / height / length / batch_size, the four tiling knobs (tile_size / overlap / temporal_size / temporal_overlap), and a final latent output for samples.

Install

Standard pack install:

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

or ComfyUI Manager → search "WanImageToVideoTiled" → Install, then restart. No extra dependencies or model downloads - it uses your existing Wan reference-subject checkpoint and VAE, plus ComfyUI's core Wan latent-format helpers.

The honest caveats

The README leaves this one unchecked - untested by the author, same as FunInpaint, Camera, and Track. Given how rarely the whole pack is discussed in the community, treat the two-negative-output behavior as "experiment and see" rather than documented behavior. And the pack-wide reminders still apply: tiling only relieves the VAE encode stage, so if the 14B denoiser is your bottleneck this won't fix it, and pair the front end with VAE Decode (Tiled) because decode is usually the slowest step of a Wan render.