ThinkSound is an open-source Any2Audio generation framework from Alibaba's FunAudioLLM team that generates and edits audio from video, text, and audio inputs, guided by step-by-step Chain-of-Thought (CoT) reasoning from multimodal large language models. Accepted to the NeurIPS 2025 main conference, it targets one of the harder problems in generative audio: producing sound that is not just plausible but semantically and temporally aligned to what is happening on screen. The project has grown into a small ecosystem, with a follow-up model, PrismAudio (ICLR 2026), living on a separate branch of the same repository. ## Reasoning-Guided Audio Generation Most video-to-audio systems map pixels to a waveform in one shot, which makes it hard to control which sounds appear and when. ThinkSound instead treats sound design as a reasoning problem: a multimodal LLM produces a Chain-of-Thought that decomposes a scene into events, and a flow-matching audio model uses that reasoning to synthesize aligned audio. This compositional approach is what lets the system claim state-of-the-art results on multiple video-to-audio (V2A) benchmarks, and it is trained with the help of AudioCoT, a large-scale CoT-annotated dataset the authors open-sourced on Hugging Face. ## A Three-Stage Interactive Workflow Rather than a single black-box pass, ThinkSound splits generation and editing into three interactive stages. First, Foley Generation produces foundational, temporally aligned soundscapes directly from video. Second, Object-Centric Refinement lets users click on or select a region of the video to refine or add sounds for a specific object. Third, Targeted Audio Editing modifies the generated audio using high-level natural-language instructions. Because all three stages share one unified foundation model, the same system covers automatic generation, precise object-level control, and text-driven editing without switching tools. ## Any2Audio Flexibility The framework is not limited to video. It accepts arbitrary modalities — video, text, audio, or combinations — and produces audio conditioned on whatever context is supplied. This makes it useful beyond Foley for film and short-form video: it can serve as a text-to-audio generator, an audio editor, or an interactive sound-design assistant. A July 2025 update lightweighted the model and optimized memory and GPU usage to support higher-throughput generation at scale, and the team ships inference scripts, a web interface, Windows batch scripts, and PyPI dependencies for cross-platform setup. ## Ecosystem and Access ThinkSound is easy to try before installing. Online demos run on Hugging Face Spaces and ModelScope, the technical report is on arXiv (2506.21448), and finetuning code is public so teams can adapt the model to their own data. The repository also hosts PrismAudio, a follow-up accepted to ICLR 2026 that pushes video-to-audio further with multi-dimensional CoT reinforcement learning, giving researchers a clear path from the NeurIPS baseline to more recent work in the same codebase. ## Trade-offs and Limitations The project's code and models are released under the Apache 2.0 License, but it bundles a fine-tuned VAE derived from Stable Audio Open that carries Stability AI's Community License, so commercial use and redistribution of that component require prior permission — an important detail for production teams. As with most research-grade audio models, high-quality generation benefits from a capable GPU, and results depend on the quality of the input video and the reasoning step. The dual-project repository layout (ThinkSound on master, PrismAudio on a branch) is powerful but can be confusing for newcomers who need to check out the correct branch. ## Who Should Use This ThinkSound is a strong fit for researchers working on video-to-audio and controllable sound generation, and for creative developers prototyping Foley, sound design, or accessibility tools where audio must track on-screen events. The interactive object-centric editing and natural-language control make it more than a demo — it is a usable pipeline for anyone who needs aligned, editable audio rather than a single fixed soundtrack.