Stack_WanPhantomSubjectToVideo

Wan Phantom solves a problem that used to mean training a character LoRA: getting a specific subject - a person, an outfit, an object - to show up consistently in generated video without ever training anything. Feed it one or more reference images and it carries that subject's appearance into a fresh generation, and by community account it does this well even on genuinely hard cases: one early tester threw "problematic faces" at it that had defeated FaceID and other identity tools, including two women talking to each other and a clothing try-on shot, and got results described as coming out "perfect," with only the model's small 1.3B parameter count as the honest caveat ("just felt that the 1.3b models just aren't very good... anxiously awaiting the 14b").

How it works. The only content input is images - plural, meaning this node accepts more than one reference at once. That matters because Phantom's behavior with two references isn't simple averaging: one tester found that "when given two faces it rather tries to mix them" rather than picking one or keeping them cleanly separate, which is worth knowing going in if your intent was two distinct characters rather than a blend. width, height, and length (default 832×480, 81 frames - the standard ~5-second Wan clip) round out the output configuration. Output is Phantom, a config bundle for a downstream Wan sampler.

Installing it. No extra dependency for this node itself: cd ComfyUI/custom_nodes && git clone https://github.com/cardenluo/ComfyUI-Apt_Preset, restart ComfyUI, or search "ComfyUI-Apt_Preset" through ComfyUI Manager. What you need beyond that is the actual Phantom checkpoint - check the model's own GitHub release page for current weights and sizes, since the 1.3B and any larger variant will have very different VRAM footprints and this pack doesn't bundle either.

Common issues. If you're feeding two reference images expecting two distinct subjects to appear separately in the output and instead getting a blended, in-between-looking result, that's the documented mixing behavior above, not a bug in this node - a single clean reference per generation is the more reliable path if identity precision on multiple simultaneous subjects matters. If output looks noticeably softer or less coherent than showcase examples you've seen online, check which model size you actually downloaded - the widely-circulated 1.3B checkpoint is meaningfully weaker on coherence than the larger 14B variant, and a lot of impressive Phantom results online are running the bigger model. And since this node has no explicit pose or motion control input, don't expect fine-grained control over how the subject moves - Phantom's job is preserving what the subject looks like, not directing what it does; pair it with a control-video node elsewhere in this pack if you need both identity preservation and motion direction at once.