Wan VACE Multi-Ref To Video

This is the main event of the Tessiiiz Wan Video Toolkit for people still on Wan 2.1. WanVaceMultiRefToVideo (displayed as "Wan VACE Multi-Ref To Video") is, in the author's own words, a cleaner replacement for ComfyUI's built-in WanVaceToVideo - same output shape, but with start/end frame pinning, positioned middle frames, multi-reference image batches, and control-video frame filtering bolted on.

The pitch is simple: you throw in a start frame, a handful of reference images pinned to specific frames, an end frame, and optionally a control video, and out the other side you get positive, negative, latent, and trim_latent - wired to your sampler exactly like any other Wan VACE graph. The idea is that pinning a few frames lets you hold a character or object across the clip instead of hoping the model keeps it together for all 81 frames.

How it works

Like its 2.2 sibling, this node builds a pinned frame timeline and then expands - via ComfyUI's GraphBuilder - into the stock WanVaceToVideo node. Concretely it:

• Allocates length frames, pre-filled with empty_frame_level gray (default 0.5).
• Stamps the control_video (filtered by frames_to_keep_in_control_video, offset by control_start_frame) at control_mask_value strength - leave that at 1.0 unless you deliberately want weaker control.
• Pins start_image (held start_frame_hold frames), each positioned_frames ref at its position (held positioned_frame_hold), and end_image at the very end (held end_frame_hold).
• Hands the timeline plus its mask to WanVaceToVideo as control_video / control_masks.

There's one compatibility constraint worth knowing before you design a workflow around it: only one primary reference frame is passed through the internal reference_image path, for Wan VACE runtime compatibility. It's the first available of start → positioned → end → extra refs. So the pinned timeline isn't a gimmick - it's the mechanism. If you want multiple references to actually anchor the clip, they have to go in positioned_frames with positions, not all dumped into extra_reference_images. That's the README's explicit guidance and it's the thing that most people get wrong on first pass.

Inputs and outputs that matter

• vae, positive, negative - required. Wire the same sources you'd use for any Wan VACE workflow.
• length - total clip length in frames (default 81, the Wan native context).
• positions_of_positioned_frames - default 20 30 50 90, space-separated 1-based frame numbers. L/END means the last frame; negatives count back from the end. Pairs one-to-one with your positioned_frames batch in order.
• frames_to_keep_in_control_video - a frame filter for the control video. 1 20 30 keeps three frames, 1:16 keeps a range. Empty = keep everything.
• inpaint_mask - optional; multiplies into the timeline mask if you want to restrict where VACE conditioning applies.

Outputs: positive, negative, latent feed your sampler as usual. trim_latent is an INT worth understanding - it's how many latent frames the reference-image prepend added. The stock WanVaceToVideo prepends the reference as extra latent frames, so after sampling you feed trim_latent into ComfyUI's TrimVideoLatent node to strip them back off. Skip that and you get a latent that doesn't match your expected frame count.

Installing and running it

Same story as the rest of this pack: it's glue, not weights. Manager search "ComfyUI-Tessiiiz-WanVideoToolkit", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Tessiiiz/ComfyUI-Tessiiiz-WanVideoToolkit

Restart ComfyUI. You still need a Wan 2.1 VACE model (14B or 1.3B) loaded through ComfyUI's normal Wan loaders - this pack ships nothing. And remember VACE is slow: 14B clips run 15-40 minutes on a 4090 without a distillation LoRA like CausVid.

Common gotchas

• Positions vs. frames mismatch - extra refs beyond your position tokens are silently dropped. Count them.
• The off-by-one - positions are 1-based.
• Don't expect extra_reference_images to do multi-ref lifting - only the first becomes the primary reference. Positioned frames are the real multi-ref path.
• Trim the latent - wire trim_latent into TrimVideoLatent, or your clip will be the right video with the wrong length.

If your whole pipeline is Wan 2.2 native nodes, use Wan22VaceMultiRefEncode instead - this one is for the 2.1 WanVaceToVideo graph shape.