Overview
NAVA (Native Audio-Visual Alignment for Generation) is Baidu ERNIE Team’s 6.3-billion-parameter model for jointly generating synchronized audio and video. Its Align-then-Fuse MMDiT architecture first aligns the audio and video modalities, then incorporates contextual conditioning during joint denoising. (arxiv.org)
Capabilities and uses
- Generates audio and video from a text prompt; it can also use a first-frame image to condition generation.
- Supports speech with reference-timbre control, including multiple speakers, and prompt-based descriptions of camera composition and motion.
- Potential uses include audiovisual content creation, dialogue scenes and research into synchronized multimodal generation. The authors report evaluations on Verse-Bench and Seed-TTS, as well as a user study; these results should be interpreted in the context of those tests. (huggingface.co)
Limitations and deployment
- The model card recommends 1280 × 704 resolution, lists audio up to 10 seconds, and documents up to two speaker reference WAV files per sample.
- It says the model was trained on dense Chinese captions and recommends rewriting short or English prompts. Single-GPU inference is supported but slower; the documented fast-generation configuration uses eight GPUs.
- Model weights are published under Apache-2.0, but the distribution also includes upstream components with their own license terms. Use reference images and voices with appropriate consent. (huggingface.co)
Release date
The research paper was submitted on May 28, 2026. This is the verified publication date of the paper, not necessarily the date on which every model artifact became available. (arxiv.org)
Pricing
As of November 16, 2026, a NAVA-specific official API price could not be verified on either Baidu’s mainland China pricing page or its international pricing page. Input and output charges, billing units, and API availability for NAVA are therefore unverified. Other models’ prices on those pages should not be applied to NAVA. Self-hosting may incur infrastructure costs. (cloud.baidu.com)