WanVacePhantomExperimental

The name is doing real work here. "Experimental" doesn't mean broken - it means this is the WanVaceAdvanced pack's kitchen sink, the node that throws two VACE control contexts and the Phantom embedding region controls at you all at once. If you found this page you're probably either looking at someone else's workflow that uses it, or you started with WanVacePhantomSimple, hit its limits, and want the dual-control version. Both are good reasons to be here.

Quick context if you're new to the pack: Wan is Alibaba's Apache 2.0 video model, VACE is its video-conditioning system (think ControlNet for video - you feed it a pose/depth control video plus a reference image of the subject), and Phantom embeddings are a separate trick for locking character identity from still images. This pack's whole job is making VACE's per-frame strength control and Phantom combination precise. The README's own summary: it gives you "fine-grained control over Vace frame strengths with reference images, control videos, and optional phantom embeddings."

What makes this one different

Compared to the plain WanVacePhantomSimple, this node adds a full second VACE channel (control_video2, control_masks2, vace_reference_2, vace_strength2, vace_ref_strength2) - handy when you want pose from one video and depth from another at the same time. And it exposes the three Phantom region knobs you don't get elsewhere:

• phantom_mask_value - the VACE mask value used for the Phantom embed region (default 1.0).
• phantom_control_value - the padded value for the VACE-embedded latents in that region (default 0.0).
• phantom_vace_strength - and this one's worth reading twice, because the author added a tooltip to stop you misreading it: it's the strength of the VACE embedding applied to the Phantom region, not the strength of the Phantom embeds themselves.

The rest of the inputs

Required are positive, negative, vae, and the frame geometry (width 832, height 480, length 81, batch_size 1). Note length is in video frames - each latent frame covers 4 video frames. The regular strength inputs (vace_strength, vace_ref_strength) accept a plain float, and with a patched model they'll also take a list of floats for per-frame control - e.g. [1.0, 1.0, 1.0, 0.5, 0.5, ...] to hold character strength high for the first few frames of a continuation, then ease off. Feed a list from this pack's own StringToFloatListRanged or Kijai's String to Float List.

Outputs are positive, negative, neg_phant_img (the extra negative you should Conditioning Combine into your main negative when using Phantom - the README says to), a latent, and trim_latent (an INT telling you how many reference frames got eaten, so downstream nodes can stay aligned).

The catch: it doesn't patch the model itself

This original-family node has no model input. To get per-frame strength lists working you must patch the model separately with VaceAdvancedModelPatch (the V2 nodes fold this in automatically). The patch is safe - untouched workflows behave identically - but forget it here and strength lists silently degrade. That's the single most common way people get confused by this node.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/drozbay/ComfyUI-WanVaceAdvanced
# restart ComfyUI

Or search "ComfyUI-WanVaceAdvanced" in ComfyUI Manager. Dependencies are light - just numpy and scipy - but you'll want Kijai's ComfyUI-KJNodes for Wan model loading and VideoHelperSuite for video I/O, plus the VACE module + Phantom 14B weights (the pre-merged InnerReflections Wan2.1_VACE_Phantom checkpoint is the easy route).

One honest warning from the community: combining Phantom and VACE is where workflows get fiddly, and people do hit weaker character consistency during inpainting with this stack. Don't blame the node first - check your CFG (Phantom needs > 1.0, 1.5–3.5 typical), your negative combining, and that your subject reference isn't cluttered. If you don't need the second control channel, WanVacePhantomSimpleV2 is the calmer default.