公司指南

Stability AI

Stability AI 是一家位於倫敦的生成式 AI 公司,成立於 2019 年,最為人知的是在 2022 年 8 月釋出 Stable Diffusion 的權重,讓任何人都能在消費級顯示卡上執行具備相當能力的文字轉圖像模型。

聆聽此指南AI 語音敘述 · ~8 min

本頁閱讀時間 9 分鐘
  1. 概述
  2. 重點摘要
  3. 深入探討
  4. 檢查 Stability 模型是否適合此專案
  5. 現實世界的實施
  6. 穩定人工智慧的未來
  7. 付諸實踐
  8. 資料來源與延伸閱讀
  9. 不斷探索
  10. 常見問題

概述

它現在為創意產業打造影像、影片、音訊與 3D 模型,其歷史亦成為開放權重授權、AI 版權法與新創經濟學的參考案例。

重點摘要

  1. Stability AI 在 2022 年 8 月釋出了 Stable Diffusion 的權重,使得具備相當能力的文字轉圖像模型能在消費級顯示卡上執行。
  2. Stable Diffusion 使用潛在擴散,對壓縮的 64×64×4 潛在空間進行去噪,而非 512×512 像素的圖像。
  3. Prem Akkaraju 於 2024 年 6 月成為 CEO,原因是 Emad Mostaque 辭職且供應商免除約 1 億美元的債務。
  4. Stability AI 社群授權對年營收低於 100 萬美元的組織提供免費商業使用。
  5. 2025 年 11 月英格蘭高等法院的裁決認為 Stable Diffusion 的權重並非侵權複製 Getty Images 的照片。

深入探討

Stability AI is a private company headquartered in London that develops generative AI models, systems that create new images, video, audio, 3D assets or text from a prompt. It was founded in 2019 by Emad Mostaque and Cyrus Hodes. Its name is tied to one model family, Stable Diffusion, a text-to-image model whose weights, the learned numbers that make up the trained network, were released for download in August 2022. That release made Stability AI one of the most visible advocates of open-weight AI, meaning models that anyone can download and run, as opposed to models available only through a company's own website or API.

Stable Diffusion is built on latent diffusion, a method published in December 2021 by Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser and Bjorn Ommer and accepted to CVPR 2022. A diffusion model learns to remove noise from an image step by step; to generate, it starts from pure noise and denoises repeatedly until an image appears. Running that process on every pixel is expensive. The latent diffusion paper instead trains an autoencoder, a network that compresses an image into a smaller representation called a latent and expands it back, and runs diffusion in that compressed space. The authors report that this greatly reduces compute compared with pixel-space diffusion while preserving detail. Text guides the process through cross-attention, a layer that lets the image representation look up information in an encoded version of the prompt.

The Stable Diffusion v1-4 model card shows the concrete numbers. Its autoencoder shrinks each side of the image by a factor of 8, so a 512 by 512 pixel image becomes a 64 by 64 grid with 4 channels. Prompts are encoded by a frozen CLIP ViT-L/14 text encoder. The checkpoint was initialized from v1-2 and fine-tuned for 225,000 steps at 512 by 512 resolution on LAION-Aesthetics v2 5+, a subset of the web-scraped LAION-2B English image and caption dataset. Training used 256 Nvidia A100 GPUs, and the card estimates the Stable Diffusion v1 models together used about 150,000 GPU hours and emitted about 11,250 kg of CO2 equivalent. The weights were released under the CreativeML OpenRAIL-M license, which permits commercial use but forbids listed harmful uses.

The company grew quickly and then hit trouble. In October 2022 it raised $101 million in a round led by Coatue and Lightspeed Venture Partners, which Wikipedia reports valued it at $1 billion. Costs outran revenue. TechCrunch, summarizing earlier reporting, said Stability was projected to make about $11 million in sales in 2023 while spending about $99 million a year on cloud computing from AWS, Google Cloud and CoreWeave, and had about $4 million in the bank in October 2023. Several Stable Diffusion researchers left; Wikipedia notes that departures from the Freiburg office preceded the founding of Black Forest Labs there in 2024. Mostaque resigned as CEO on March 23, 2024. On June 25, 2024, Prem Akkaraju, former CEO of the visual effects company Weta Digital, became CEO, Sean Parker became executive chairman, and suppliers agreed to forgive about $100 million of debt and release the company from about $300 million in future obligations.

Under new leadership the product line moved toward professional creative work. Stable Diffusion 3.5 was released on October 22, 2024, in three variants: Large with 8.1 billion parameters for 1-megapixel images, Large Turbo, a distilled version that generates in 4 steps, and Medium with 2.5 billion parameters, added October 29, which the company says needs 9.9 GB of graphics memory excluding text encoders. These models are built from transformer blocks, rather than the convolutional U-Net used in the original latent diffusion work. The company also lists video, audio and 3D models. On August 25, 2026, it announced a $76 million Series B with new investors including Electronic Arts, Sony Music Group, Universal Music Group, Warner Music Group and AMD Ventures, bringing funding since Akkaraju's arrival to $232 million including convertible notes.

Compared with closed image generators that are offered only as hosted services, Stability's open-weight approach lets users run models privately, modify them and build on them, and it created a large ecosystem of community fine-tunes and tools. The trade-off is control: once weights are public, the company cannot enforce safety filters on every copy, and it earns no revenue from most use. Its licensing has changed accordingly, from the permissive OpenRAIL-M license of 2022 to the Community License, which is free for non-commercial use and for organizations under $1 million in annual revenue, with larger organizations required to buy an enterprise license.

The largest open questions are legal. Artists filed a US lawsuit against Stability AI, Midjourney and DeviantArt in January 2023, and Getty Images sued in both the United States and England, alleging that millions of its photos were used for training. In the English case, Getty Images v Stability AI [2025] EWHC 2863 (Ch), Getty dropped its primary claims about training during the trial, and the court noted that training had taken place outside the UK. The court held that model weights can in principle be an article under UK copyright law, but that Stable Diffusion's weights were not an infringing copy because they never stored Getty's images. Getty partly succeeded on trademark claims over generated images showing its watermark, and permission to appeal the copyright point has been granted.

The model cards also list technical limitations: version 1 models struggle with legible text, complex compositions and faces, work best with English prompts, lose detail through the lossy autoencoder, and can reproduce images that were duplicated many times in training data. In practice, running a Stability model requires a graphics card with roughly 7 to 10 GB of memory for the smaller models, more for the largest, or use of the company's paid API and partner clouds, plus a check of the license that applies to the specific model version.

04工作範例

檢查 Stability 模型是否適合此專案

  1. 選擇精確的模型版本,例如 Stable Diffusion 3.5 Medium,並開啟其官方發布說明或模型卡。

  2. 閱讀授權:在社群授權下,注意貴組織的年度營收是否低於 100 萬美元。

  3. 將記憶體需求與硬體做比較;公司列出 3.5 Medium 需要 9.9 GB 顯示記憶體,排除文字編碼器。

  4. 閱讀模型卡的訓練資料與限制部分,並注意任何相關事項,例如版本 1 模型的文字渲染較弱。

  5. 以模型原生解析度產生一小批測試樣本,並檢查是否出現列出的失敗模式。

  6. 記錄所使用的版本、授權與設定,以便日後追蹤輸出。

它顯示了什麼

學習者以文件化決策結束,確認特定 Stability 模型是否能合法且實際使用,並依官方授權條款、硬體規格與已說明的限制進行檢核,而非僅依賴對 Stable Diffusion 的一般說法。

現實世界的實施

本地影像生成:2022 年 8 月的 Stable Diffusion 版本表示模型大約需要 6.9 GB 的顯示記憶體,讓業餘者與研究人員可以在自己的電腦上執行,而不必使用付費雲端服務。

微調衍生模型:由於 Stable Diffusion 的權重是公開的,開發者會針對特定風格或產品進行微調,而 3.5 版在其 transformer 區塊中加入了查詢‑鍵正規化,部分目的是讓微調更簡單。

小型企業商業使用:根據 Stability AI 社群授權,年營收低於 100 萬美元的組織可免費商業使用 Stable Diffusion 3.5,較大的公司則需取得企業授權。

音樂製作:2026 年 Stability AI 發布了 Stable Audio 3.0,這是一系列開放權重的音樂模型,據稱使用完全授權的資料訓練,可透過數位音訊工作站的插件取得。

法律先例:Getty Images 訴 Stability AI 案件,由英格蘭及威爾斯高等法院於 2025 年 11 月裁決,是英國關於訓練模型權重是否構成侵權影像複製的具指標性判決。

穩定人工智慧的未來

Stability AI's 2026 direction is toward licensed and professional creative tools. Its August 2026 Series B announcement named record labels and a game publisher as investors, said Stable Audio 3.0 was trained on fully licensed data, and said the money would fund a creative production product suite, applied research and professional services. Whether licensed-data models can match the quality of models trained on large web scrapes is an open question for the whole field. Legal outcomes remain unsettled. The November 2025 English High Court ruling decided only that Stable Diffusion's weights were not an infringing copy under UK law, not whether training itself infringes, and permission to appeal has been granted. In the United States, Getty's separate case and the artists' class action were still in progress in 2026, with Wikipedia reporting that a federal court allowed most of Getty's US claims to proceed to discovery in April 2026. Open-weight release itself faces pressure from several directions: the cost of training competitive models, the difficulty of enforcing license terms and safety limits on downloaded weights, and competition from both closed models and other open-weight developers, including Black Forest Labs, founded by former Stability researchers. How Stability balances open releases with paid enterprise licensing is likely to keep changing with each model generation.

付諸實踐

為什麼這很重要

  • 供應商策略. 供應商路線圖會影響您的團隊接下來可以建立的功能。
  • 成本與預算. 商業條款和部署選項會影響長期成本和風險。
  • 風險與安全. 公司激勵措施塑造了產品預設、安全態勢和開放性。

留意

  • 發佈公告可能會超過實際生產工作流程的穩定性。
  • API 定價或政策轉變可能會在一夜之間打破假設。
  • 單一供應商依賴性增加了鎖定和遷移成本。

後續步驟

  1. 使用您自己的任務和資料集評估提供者。
  2. 在整合之前查看隱私、安全和法律條款。
  3. 維護跨模型或供應商的後備計劃。
  4. 監控發行說明,以便路線圖的變更不會讓團隊感到意外。

資料來源與延伸閱讀

  1. arXiv (CVPR 2022)使用潛在擴散模型的高解析度影像合成
  2. Hugging FaceCompVis/stable-diffusion-v1-4 模型卡
  3. Stability AIStable Diffusion 公開發佈
  4. Stability AI全新推出 Stable Diffusion 3.5
  5. TechCrunchStability AI 獲得 Sean Parker 與 Greycroft 的救援資金
  6. Stability AI娛樂產業的重量級人物在最新一輪融資中支持 Stability AI
  7. AIPPIGetty Images 訴 Stability AI:英國高等法院在里程碑式生成式 AI 判決中認定未構成次要著作權侵害
  8. WikipediaStability AI

不斷探索

免費電子報

獲取每日 AI 簡報

每個工作日上午提供三則經過驗證的 AI 新聞,以淺顯易懂的英文撰寫。永久免費,無廣告。

每個工作日一封電子郵件。一鍵取消訂閱。我們絕不出售或分享您的地址。

測試一下自己

參加 Stability AI 測驗

每個答案皆有即時回饋,通過課程後可獲得具備可驗證 ID 的可分享證書。

開始測驗

支持免費的 AI 教育。AI Understanding 是一個 501(c)(3) 非營利組織——絕無廣告,絕無付費牆。 捐款

常見問題

什麼是Stability AI?

Stability AI 是一家位於倫敦的生成式 AI 公司,成立於 2019 年,最著名的是於 2022 年 8 月發布 Stable Diffusion 的權重,讓任何人都能在消費級顯示卡上運行一個功能強大的文字到圖像模型。它現在為創意產業構建圖像、影片、音訊和 3D 模型,其歷史成為開放權重授權、AI 版權法和創業經濟學的參考案例。

Stability AI 在 2022 年發布的最出名的產品是什麼?

Stability AI 支援 2022 年 8 月公開發布 Stable Diffusion,其可下載權重使其成為開源生成人工智慧的里程碑。

Stability AI 在 2022 年 Stable Diffusion 發布時公開了什麼?

Stability AI 描述了可公開下載的 Stable Diffusion 權重,並受發布條款限制。這將主張限定於其自身發布,並不比較其他公司的存取條款。

從技術上講,穩定擴散在哪裡執行其去雜訊過程?

它是一個潛在擴散模型:自動編碼器將影像壓縮到更小的潛在空間中,U-Net 在其中進行去噪,從而大幅削減計算量。

根據 Stable Diffusion v1-4 模型卡,該檢查點使用了哪些訓練資料?

v1-4 模型卡表示該模型是從 v1-2 初始化,然後在 LAION-Aesthetics v2 5+ 子集上進行 225,000 步的微調。卡片還列出了 LAION-2B(en) 及其子集作為訓練資料。這是版本特定的;不要將 LAION-5B 視為一個未經過濾的集合,並在每個檢查點上相同使用。

除了影像之外,Stability AI 還為哪種模式建立模型?

Stability AI 已發布 Stable Audio 文字到音訊模型,包括 Stable Audio Open。此項僅描述音訊模態;對公司語言、3D 或影片模型的說明需要各自的來源。