Why this is a different model
T2V/I2V DiTs are trained to denoise RGB (or VAE) latents. A V2A model denoises audio latents conditioned on video features (CLIP/Sync), captions, and often chain-of-thought. Mixing those heads into one 2.2 open ckpt was not part of the July 2025 release.
ThinkSound in this family of tools
- Paper: Chain-of-Thought Reasoning in MLLMs for Audio Generation and Editing (arXiv:2506.21448, NeurIPS 2025).
- Three stages: foley from video, object-centric refine (click/region), language-guided edit.
- Backbone: MM-DiT with VideoLLaMA 2 CoT; weights on Hugging Face FunAudioLLM/ThinkSound.
Practical mux
Keep the Wan clip’s frame rate and duration. Generate audio at a matching length, then ffmpeg -c:v copy -c:a aac. If you need lip-sync speech, go to S2V instead of V2A — V2A will not rebuild a specific actor’s mouth from a wav you already have.
FAQ
- Can TI2V-5B output AAC?
- No. Decode is video frames. Any soundtrack is a second model or a later closed Wan API.
- Is PrismAudio the same?
- PrismAudio is a follow-up V2A line (CoT-RL) from the ThinkSound authors. Same problem family, different checkpoint.
Primary sources
Checked against Wan-Video/Wan2.2 README · Aug 28, 2026
