Baidu ERNIE team's 6.3B joint audio-video generation model with native audio-visual alignment and timbre-controlled speech.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
NAVA (Native Audio-Visual Alignment) is a research framework from Baidu's ERNIE team for joint audio-video generation. It uses an Align-then-Fuse MMDiT architecture: Hierarchical Alignment Layers first establish audio-video correspondence in a dedicated alignment space, then Unified Fusion Layers perform context-conditioned denoising of audio and video in a single trajectory. The project page credits Wan2.2-TI2V-5B, LTX-Video, ReDimNet, Qwen3 and Ovi as building blocks.\n\nThe model adds Timbre-in-Context Conditioning, which binds a reference voice's timbre to the corresponding speech spans, enabling controllable speech generation. It supports text-to-audio-video (T2AV) and text-plus-image-to-audio-video (TI2AV) generation. On Verse-Bench the authors report the best audio-visual synchronization and video quality among compared open models (Ovi 1.1, MOVA, daVinci, LTX 2.3) with 6.3B parameters, and a human preference study shows wins on audio-visual alignment. These are the authors' own results.\n\nA YouTube creator summary describes it as producing 720p video with synchronized speech, with a roughly 25GB checkpoint and a 7GB FP8 version, and as open-source. The fetched project page did not link code, weights or a paper, so release details, license and download locations could not be verified here.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Video generation framework from Baidu's ERNIE team that adds natively aligned audio to Wan 2.2, producing 720p video with synchronized speech. Reportedly beats LTX 2.3 despite a far smaller size; open-source, 25GB with a 7GB FP8 version.
Closest neighbours by meaning, not just by category.
Model · Kandinsky Lab
Open-source diffusion models for synchronized 5-second video and audio generation, with a Full HD super-resolution add-on.
Model · Meta
Meta research model that generates audio directly in raw waveform space, without a VAE, for video-to-audio synthesis.
Model · ByteDance
Research zero-shot TTS model that generates speech by modeling raw waveforms directly with flow-matching DiT.
Model · Google
Google DeepMind's video generation model producing clips with natively generated audio, dialogue and sound effects.
Model
Research model for editing video and audio jointly from a text instruction, e.g. changing speech with lip sync or voice.
Model · Pika Labs
Pika's open-weight video model post-trained on MiniMax H3, generating synchronized video and audio up to 4K.
