WanVideo Add Bindweave Embeds

BindWeave is ByteDance's subject-consistent video model - the problem it targets is keeping a specific subject (or several) looking like themselves across a generated clip, for both single- and multi-subject prompts. Its trick is cross-modal: it couples a multimodal LLM (Qwen-VL) with the video DiT, so the model actually understands which subject in your prompt maps to which reference. This node folds BindWeave's conditioning into a WanVideoWrapper graph. You give it your image embeds plus reference latents (and, optionally, the Qwen-VL embeds that carry the semantic subject binding), and it produces augmented image embeds that steer the generation toward your references.

It's one of the more capable identity/subject tools in this pack, and the multi-subject support is what sets it apart from a plain single-face lock.

How it works

BindWeave binds subjects to a language understanding of the prompt. The reference latents give it the visual "who," and the optional Qwen-VL embeds give it the semantic "which subject is which" - which is how it handles multiple people or objects without smearing them together. This node combines all that into your WANVIDIMAGE_EMBEDS, and helpfully spits out preview outputs so you can sanity-check what it built before spending a generation on it.

The inputs and outputs that matter

• embeds (WANVIDIMAGE_EMBEDS) - your image embeds heading to the sampler.
• reference_latents (LATENT) - the encoded reference subject(s). This is the visual identity being carried through.
• ref_masks (MASK, optional) - which region each reference applies to, the key to clean multi-subject binding.
• qwenvl_embeds_pos / qwenvl_embeds_neg (QWENVL_EMBEDS, optional) - the multimodal-LLM conditioning that ties prompt subjects to references. This is BindWeave's cross-modal secret sauce; wire it in for the full behavior.

Outputs: the augmented image_embeds (WANVIDIMAGE_EMBEDS) into your sampler, plus image_embed_preview (LATENT) and mask_preview (MASK) so you can verify the binding looks right.

How to install it

ComfyUI Manager - search ComfyUI-WanVideoWrapper, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper
pip install -r ComfyUI-WanVideoWrapper/requirements.txt

then restart. BindWeave needs its model weights (ByteDance's BindWeave release, with Kijai's ComfyUI-ready versions on WanVideo_comfy) plus a Qwen-VL model to produce the QWENVL_EMBEDS - that's a meaningful extra download and VRAM cost on top of the Wan base.

Common issues & troubleshooting

Multiple subjects blend into one. Without ref_masks (and ideally the Qwen-VL embeds), the model has no clean way to keep subjects apart. Multi-subject is exactly where you need the masks and the semantic binding wired up - a bare reference-latent connection tends toward mush.

Consistency is weaker than you hoped. BindWeave is strong but it's a subject-consistency framework, not a guarantee, and it's built on the aging Wan base whose limits still apply. Clean, distinct reference images help; muddy or similar-looking references make the binding job harder.

Check the previews before you commit. That's what image_embed_preview and mask_preview are for. If the mask preview shows your regions in the wrong place, fix the masks before you spend twenty-plus minutes rendering the wrong binding.