WanVideo Ovi CFG

Ovi is the mode inside WanVideoWrapper that generates a video and its own audio together - actual speech and sound synced to the clip, from a single text or image prompt, built on the Wan 2.2 5B architecture. It's a genuinely different beast from normal Wan: you're not just guiding pixels, you're guiding a soundtrack too, and those two streams don't necessarily want the same amount of guidance strength. WanVideoOviCFG is the small node that lets you tune the audio side separately from the video side.

What it actually does

It takes the WANVIDEOTEXTEMBEDS your normal text-encode node already produced and adds two things to that bundle: an ovi_audio_cfg value, and optionally a separate negative embed just for the audio stream. It doesn't re-encode anything or touch your video prompt - it's attaching audio-specific guidance onto conditioning that already exists, so the sampler can apply one CFG to what you see and a different one to what you hear.

The inputs and output that matter

• original_text_embeds (required) - the WANVIDEOTEXTEMBEDS from your normal text encoder. This node modifies and passes it through.
• ovi_audio_cfg (default 3, range 0–100) - the guidance strength for the audio generation specifically. This is the knob you're here for.
• ovi_negative_text_embeds (optional) - a negative prompt dedicated to the audio track, distinct from whatever negative you're already running for the video. Build it with WanVideoTextEncodeSingle rather than the combined text-encode node, since you just want one standalone embed here, not a pos/neg pair.

Output is a single text_embeds (WANVIDEOTEXTEMBEDS) that goes wherever your original embeds were going - into the sampler.

How to install it

Ships with the wrapper, no separate download for this node. 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

Ovi mode itself needs its own model files beyond the standard Wan setup: a dedicated diffusion model, VAE, and vocoder for the audio side, plus the umT5 text encoder. People running it pull these from Kijai's WanVideo_comfy_fp8_scaled repo on HuggingFace (look under TI2V/Ovi) - grab the example workflow from the pack's example_workflows/ folder and let it tell you exactly which files go where.

Common issues & troubleshooting

It's slow. This isn't a fast mode. Community reports put a 10-second 480p clip at around 30 minutes on a 4060 Ti 16GB, even with TorchCompile and SageAttention running. Budget for it, or treat it as an overnight-batch tool rather than something you iterate on live.

The audio sounds robotic. This is the honest, consistently-reported limitation, not a settings problem - Ovi's speech generation lags dedicated TTS models in emotion and naturalness. ovi_audio_cfg can sharpen adherence to your text, but it won't fix flatness on its own.

The last word of the audio gets clipped. A known quirk people work around by padding their prompt with an extra throwaway word at the end, so the real content doesn't get cut off.

Video looks overblown or blown-out. That's usually the main video CFG, not ovi_audio_cfg - try dropping it (some report needing to go as low as ~1.7) rather than adjusting the audio scale.

LoRAs don't transfer well. Ovi runs on the 5B backbone, and standard 14B Wan LoRAs generally don't apply. A common workaround is treating Ovi's output as a base pass, then re-driving it through WanAnimate or InfiniteTalk where your 14B LoRAs actually work.