NAVA¶

Development line: project:nava · thread nava-development
Last event: 2026-06-03 · 1 dated since 2026-06-03 · Researched: 2026-09-06 · confidence: medium

What it is¶

NAVA is a local joint audio-video model for text-to-audio-video, first-frame image continuation, audio-only generation, and reference-timbre speech.

• Generates video and stereo scene/speech audio together instead of attaching audio after video generation.
• Uses an Align-then-Fuse MMDiT; reference WAVs map to ordered <S>…<E> speech spans.
• 1280×704 at 37 frames/24 fps runs for about six seconds; full-resident inference is documented at 80 GB VRAM, while the FP8 path is about 18 GB.
• Use the maintained local code or ComfyUI path; the official hosted Space was paused when checked.

Development line¶

• 2026-06-03 — NAVA public project resources were linked. On 2026-06-03, NAVA linked its project page, GitHub repository, and Hugging Face page to direct readers to documentation, source code, and model resources.

What changed¶

• 2026-05-28 — NAVA paper v1 described a 6.3B Align-then-Fuse MMDiT for joint audio-video generation.
• 2026-05-29 — NAVA reached its official initial model release.
• 2026-05-30 — NAVA uploaded NAVA.safetensors and removed the Hub copy of legacy NAVA.ckpt.
• 2026-06-01 — NAVA added the NAVA_fp8.safetensors lower-VRAM checkpoint.
• 2026-06-03 — NAVA uploaded English and Simplified-Chinese model-card documentation; GitHub also updated the README and optimized inference.
• 2026-06-05 — NAVA added FP8 support for ComfyUI.
• 2026-06-08 — NAVA optimized and cleaned up its ComfyUI implementation.
• 2026-06-15 — NAVA fixed FP8 prompt-rewriter desynchronization and added a vision-language rewrite workflow.
• 2026-06-23 — NAVA updated its Hugging Face Space and Gradio implementation.
• 2026-06-30 — NAVA updated VerseBench inference scripts.

How to use this¶

1. Treat the official hosted demo as unavailable for now: its Space was paused when checked; use a local workflow instead. — https://huggingface.co/spaces/baidu/NAVA
2. Clone the official code, install CUDA-matched PyTorch first, then install NAVA and FlashAttention in the documented order. — https://github.com/ernie-research/NAVA
3. Download the maintained weights from the current baidu/NAVA model page; use NAVA.safetensors for bf16 or NAVA_fp8.safetensors for the lower-VRAM path. — https://huggingface.co/baidu/NAVA
4. Put each job in JSONL: use prompt for text-to-audio-video, add image_path for first-frame conditioning, and add up to two spk_wavs for timbre-controlled speech; then choose the matching inference script. — https://github.com/ernie-research/NAVA
5. For a node workflow, install the official ComfyUI nodes, select NAVA_fp8.safetensors, and load one of the supplied example graphs. — https://github.com/ernie-research/NAVA/blob/main/comfyui_nava/README.md

Best practices¶

• Keep the CUDA-matched PyTorch install, editable NAVA install, and flash-attn --no-build-isolation as separate steps; do not force-reinstall the requirements file. — https://github.com/ernie-research/NAVA
• Rewrite short or English prompts before inference, and preserve every <S>…<E> speech span unchanged. — https://github.com/ernie-research/NAVA
• Keep reference WAV order aligned to speech-span order; for ComfyUI, enable T5 offload first and group offload below 48 GB VRAM. — https://github.com/ernie-research/NAVA/blob/main/comfyui_nava/README.md
• Use FP8 to reduce VRAM, not to expect faster matrix computation; it dequantizes to bf16 at compute time. — https://github.com/ernie-research/NAVA
• Use consent-approved face and voice references, label generated material as synthetic, and apply provenance or watermarking before redistribution. — https://huggingface.co/baidu/NAVA
• Default to the current safetensors checkpoints; the Hub deleted NAVA.ckpt, although local code retains checkpoint compatibility. — https://huggingface.co/baidu/NAVA/commits/main

Superseded by this¶

• 2026-05-30 — Hub download guidance requiring NAVA.ckpt is obsolete: the official Hub removed that artifact and provides safetensors checkpoints; local .ckpt compatibility remains.
• 2026-06-05 — Treating NAVA as an 8-GPU script-only workflow is outdated: the official project added an FP8 ComfyUI path.
• 2026-06-15 — Pre-fix FP8 prompt-rewrite guidance is outdated for image-conditioned work: the official code fixed rewriter desynchronization and added VL rewrite.

Still unknown¶

• The exact accompanying text of the 2026-06-03 entry is unavailable, so we treat that date as a same-day documentation and inference-update milestone, not an inferred launch announcement.
• No semantic release tags or GitHub Releases map commit-level changes to versioned NAVA releases.
• The official Space was paused when checked, so a successful current hosted-inference run was not verified.
• Official documentation is internally inconsistent: current code defaults to safetensors while parts of the model card still show deleted NAVA.ckpt paths.
• No later first-party code or model-history entry was found after 2026-06-30; this does not prove that no update exists outside the checked official histories.

Sources¶

Agent brief¶

• Subject: project:nava, thread nava-development, 1 dated events 2026-06-03 → 2026-06-03.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.