What Is Stable Audio 3.0?Stable Audio 3.0 是什么？
Stable Audio 3.0 is Stability AI's four-model family for AI music and sound effects, released May 20, 2026 as a fundamental redesign over Stable Audio 2.0 (September 2024), built on a semantic-acoustic autoencoder (SAME). Three models ship with open weights under the Stability AI Community License: Small SFX (459M parameters, runs on a MacBook CPU) for sound effects, Small (0.6B) for full music on consumer hardware (up to about 2 minutes), and Medium (2B, GPU) for longer, more musical tracks up to 6 minutes 20 seconds of stereo 44.1 kHz audio. The fourth, Large (2.7B), offers the highest quality but is available only through Stability's API or enterprise self-hosting — its weights are not public, and no timeline for opening them has been announced.Stable Audio 3.0 是 Stability AI 的 4 模型家族，做 AI 音乐与音效，2026 年 5 月 20 日发布，相对 Stable Audio 2.0（2024 年 9 月）是架构重做，基于语义-声学自编码器（SAME）。其中 3 个以开权重形式在 Stability AI Community License 下发布：Small SFX（459M 参数，MacBook CPU 可跑）做音效，Small（0.6B）在消费级硬件上做完整音乐（约 2 分钟），Medium（2B，需 GPU）做更长、音乐性更强的曲目，最长 6 分 20 秒立体声 44.1kHz 音频。第 4 个 Large（2.7B）质量最高，但只走 Stability API 或企业自托管——权重不公开，也未宣布开放时间表。
The training set is the selling point: about 1.27 million recordings — roughly 806,000 licensed from AudioSparx and about 472,000 Creative Commons recordings from Freesound, screened to remove unlicensed music. That is why Stable Audio 3.0 is pitched as the first foundation music model engineered for commercial use while rivals face active copyright litigation. Practical extras: LoRA fine-tuning for Small and Medium (with Apple Silicon workflows), audio inpainting and multi-segment editing, and ComfyUI support on launch day. Medium needs an NVIDIA GPU with CUDA and Flash Attention 2 (about 5–6.5 GB VRAM depending on length). Prices per the official site.训练集是它的卖点：约 127 万条录音——约 80.6 万条来自 AudioSparx 的授权录音，约 47.2 万条来自 Freesound 的 CC 录音，且筛掉了未授权音乐。这就是它被定位为“为商用而生的第一个基础音乐模型”的原因——对手正面临版权诉讼。实用附加项：LoRA 微调（Small 与 Medium，含 Apple Silicon 工作流）、音频内补（inpainting）与多段编辑、发布当日即支持 ComfyUI。Medium 需要 NVIDIA GPU + CUDA 与 Flash Attention 2（按长度约 5–6.5GB 显存）。价格以官网为准。
Stable Audio 3.0 is the strongest open-camp answer to the AI music copyright problem: downloadable weights, no per-track meter, prompts that never leave your machine, and a training set you can actually describe to a client. That combination matters more than raw quality for video creators, podcasters and indie game developers who need beds they can ship. The honest limits: it is not a vocal tool (no singing, no voice cloning), Large — the best quality — is API-only, generated music still needs editing for transitions and endings, and the $1M license threshold means a growing studio must re-check its rights before scaling. Verdict: if your work is instrumental beds, SFX and ambient music, start on Small or Medium locally and keep the license file with the project; if you need finished songs with vocals, use a vocal platform and treat Stable Audio as the bed layer underneath.
Stable Audio 3.0 是 AI 音乐版权问题最强的开放阵营答案：可下载的权重、没有按曲计费的表、从不离机的提示词，以及能向客户讲清楚的训练集。对需要“能交货的垫乐”的视频创作者、播客与独立游戏开发者来说，这套组合比单纯的音质更重要。真实局限：它不是人声工具（不唱歌、不克隆声音），质量最好的 Large 只走 API，生成的音乐在过渡与结尾上仍需剪辑，而 100 万美元的授权门槛意味着成长中的工作室在扩大规模前必须重查权利。结论：如果你做的是器乐垫乐、音效与氛围音乐，就在本地从 Small 或 Medium 起手，并把授权文件随项目存档；如果你要带人声的成品歌，就用人声平台，把 Stable Audio 当下面的垫乐层。
— 中国吧 AI Tools Hub Editorial Team
Who Should Use Stable Audio 3.0?谁适合使用 Stable Audio 3.0？
✓ Good Fit For适合以下场景
- Video creators and podcasters needing shippable music beds需要能交货的垫乐的视频创作者与播客
- Indie game developers needing SFX and ambient loops offline需要离线音效与氛围循环的独立游戏开发者
- Teams wanting local, private generation with no per-track fees想要本地、私密、无按曲计费生成的团队
✕ Not Ideal For不适合以下场景
- Songs with vocals, lyrics or voice cloning — different model class带人声、歌词或声音克隆的歌——那是另一类模型
- Studios above $1M revenue assuming free commercial use年收入超 100 万美元还默认免费商用的工作室
- Users with no GPU who need the top Large quality locally没有 GPU、又想在本地跑最高 Large 质量的人
Getting Started with Stable Audio 3.0Stable Audio 3.0 快速开始
1. Read the current Stability AI Community License first — check the $1M revenue threshold for your org1. 先读最新 Stability AI Community License——核对你所在机构的 100 万美元年收入门槛2. Pull Small (CPU) or Medium (GPU) weights from Hugging Face and run stable-audio-tools or ComfyUI2. 从 Hugging Face 拉 Small（CPU）或 Medium（GPU）权重，用 stable-audio-tools 或 ComfyUI 跑起来3. Generate a short bed, then edit transitions and endings — do not ship raw output to a client3. 先生成一段短垫乐，再剪过渡与结尾——不要把原始输出直接交客户Key Features核心功能
- Up to 6 Minutes 20 Seconds最长 6 分 20 秒 — Medium and Large generate over six minutes of stereo 44.1 kHz audio — full beds, not just clips.Medium 与 Large 生成 6 分钟以上的立体声 44.1kHz 音频——是完整的垫乐，不只是片段。
- Licensed-Only Training Data仅授权数据训练 — ~1.27M recordings: ~806K AudioSparx licensed + ~472K Freesound CC, screened for unlicensed music.约 127 万条录音：约 80.6 万条 AudioSparx 授权 + 约 47.2 万条 Freesound CC，且筛掉未授权音乐。
- Open Weights, Local Runs开权重、本地跑 — Small SFX (459M) on CPU, Small (0.6B) on consumer hardware, Medium (2B) on GPU — no per-track fees, prompts stay offline.Small SFX（459M）跑 CPU，Small（0.6B）跑消费级硬件，Medium（2B）跑 GPU——无按曲计费，提示词不离机。
- Inpainting & LoRA Fine-Tuning内补与 LoRA 微调 — Edit sections mid-track, continue causally, and fine-tune Small/Medium on your own audio (with documented rights).曲中改段、因果续写，并可在自有音频上微调 Small/Medium（需有 documented 权利）。
- ComfyUI Day-0 SupportComfyUI 首日支持 — ComfyUI supported Small and Medium on launch day; stable-audio-tools on GitHub is the reference runner.ComfyUI 发布当日即支持 Small 与 Medium；GitHub 上的 stable-audio-tools 是参考运行器。
Pros & Cons优缺点
✓ Pros优点
- Licensed-only training data — the cleanest legal footing in AI music仅授权数据训练——AI 音乐里最干净的法律立足点
- Open weights: no subscription, no per-track meter, private prompts开权重：无订阅、无按曲计费、提示词私密
- 6m20s stereo output with inpainting and LoRA workflows6 分 20 秒立体声输出，带内补与 LoRA 工作流
- CPU tier exists (Small SFX) for lightweight SFX work有 CPU 档（Small SFX）做轻量音效
✕ Cons缺点
- Large (2.7B) weights not public — best quality is API/enterprise onlyLarge（2.7B）权重不公开——最好质量只走 API/企业
- No vocals, lyrics or voice cloning — music and SFX only无人声、歌词或声音克隆——只有音乐与音效
- Community License is revocable with a $1M revenue cutoffCommunity License 可撤销、且有 100 万美元收入门槛
- Raw output still needs editing (transitions, repetition, endings)原始输出仍需剪辑（过渡、重复、结尾）
Use Cases应用场景
Popular ways people use Stable Audio 3.0.人们使用 Stable Audio 3.0 的常见场景：
🎬 Video Music Beds视频垫乐
Generate matching beds across cuts for AI-video renders, without a stock-library subscription.为 AI 视频成片跨镜头生成风格统一的垫乐，不用订素材库。
🎮 Game SFX & Ambience游戏音效与氛围
Small SFX on CPU for quick effects; Medium for ambient loops that run under gameplay.CPU 上的 Small SFX 做快速音效；Medium 做游戏底层的氛围循环。
🎧 Podcast Scoring播客配乐
Intros, outros and beds in a consistent style, generated and edited locally.风格统一的片头、片尾与垫乐，本地生成、本地剪辑。
🏢 Client-Safe Commercial Audio可交客户的商用音频
Under the $1M threshold, outputs commercialise with a describable licensed-data provenance — keep the license file with the project.在 100 万美元门槛内，输出可商用且训练数据来源讲得清——把授权文件随项目存档。
Similar AI Tools相似 AI 工具
If you need finished songs with vocals, see Mureka (lyrics-first songs, stems, voice cloning on Pro) — Stable Audio 3.0 is the bed layer underneath, not the singer. For the vocal-first flagship comparison, Suno remains the reference point, with its own litigation exposure to weigh.如果要带人声的成品歌，请看 Mureka（歌词优先的歌、分轨、Pro 档声音克隆）——Stable Audio 3.0 是下面的垫乐层，不是歌手。人声优先的旗舰参照仍是 Suno，但要权衡它自身的诉讼风险。