WanVideo TextEncodeSingle

The wrapper's main text-encode node, WanVideoTextEncode, is a two-for-one: it takes a positive and negative prompt and hands back one bundled WANVIDEOTEXTEMBEDS object, plus a |-separated prompt-travel trick on top. Most of the time that's exactly what you want. But some slots in the graph don't want a pos/neg pair - they want one standalone embed, and only one. WanVideoTextEncodeSingle is the plain version: one string in, one embeds object out, no bundling.

The clearest case where you need this: WanVideoOviCFG's ovi_negative_text_embeds input. Ovi's audio-video mode wants its own dedicated negative prompt for the audio track, separate from your main video negative - and that slot expects a single embed, not a positive-plus-negative pack. This node is how you build it.

How it works

Your prompt string goes through the umT5 text encoder - the same one WanVideoTextEncode uses - and comes out as a WANVIDEOTEXTEMBEDS object holding just that one encoded prompt. No pairing, no travel syntax, no assumptions about what the embed is "for." It's a raw building block other nodes then consume however they need to.

The inputs and output that matter

• prompt (multiline string, required) - the text to encode. Whatever you'd type into the positive or negative box on the combined node, minus the pairing.
• t5 (optional) - the loaded WANTEXTENCODER. Required in practice; nothing gets encoded without it.
• model_to_offload (optional) - hand it your loaded WANVIDEOMODEL and this node will offload it from VRAM while the T5 encoder runs, then it comes back after. Standard VRAM-management pattern in this wrapper - useful if you're tight on memory and don't want the diffusion model and the text encoder both resident at once.
• force_offload (default true) - whether that offload actually happens.
• use_disk_cache (default false) - caches the encoded result to disk so re-running the same prompt string skips the encode step. Handy if you're iterating on other settings with a fixed prompt.
• device (gpu or cpu, default gpu) - where the encode runs. CPU is slower but frees VRAM during the encode, which only happens once per prompt anyway.

Output is a single text_embeds (WANVIDEOTEXTEMBEDS) - feed it directly into whatever single-embed slot needs it.

How to install it

Ships with the pack. 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. You'll need the umT5-XXL encoder in models/LLM for the t5 input to have anything to load - same requirement as the main text-encode node.

Common issues & troubleshooting

You don't actually need this node. If you're just running a normal T2V or I2V graph, WanVideoTextEncode already gives you positive and negative in one shot - don't add complexity you don't need. Reach for the single-prompt version only when a specific slot (Ovi's audio negative, or a manual embed-combine setup) calls for one standalone embed.

Wired into the wrong slot. Because the output type (WANVIDEOTEXTEMBEDS) is identical to what the combined encoder produces, it's easy to plug a single-prompt embed somewhere expecting a full pos/neg pair. If conditioning seems to be missing a side entirely, check whether you actually needed the bundled node instead.

t5 not connected. Same failure as the combined node - no encoder, no encode. Make sure your T5 loader is actually wired in before troubleshooting anything downstream.