Runway Solaris評測：無程式碼介面世界模型 | Runway Solaris Review: No-Code UI, Frame by Frame

9月 01, 2026

By Kit 小克 | AI Tool Observer | 2026-09-02

🇹🇼 Runway Solaris評測：無程式碼介面世界模型

Runway 在 8 月 31 日發表 Solaris，號稱是全球第一個「Interface World Model（介面世界模型）」——不寫任何 HTML、CSS 或 JavaScript，直接像生成影片一樣，一格一格即時畫出可互動的使用者介面。這是 Runway Solaris 目前在 AI 圈討論度最高的話題，因為它跟過去所有「AI 生成網頁」工具的做法完全不同。

Interface World Model 到底在做什麼

市面上像 v0、Bolt、Lovable 這類工具，本質上是叫 LLM 寫出一份 React 或 HTML 元件樹，再由瀏覽器渲染成畫面。Solaris 反其道而行：它建立在 Runway 自家的 Gen-4.5 影片生成模型上，把「使用者的點擊、拖曳、打字」當成影片的下一幀輸入，直接輸出畫面本身，而不是輸出程式碼。換句話說，你看到的按鈕、輸入框、動畫，全部是模型即時「畫」出來的影片畫面，不是真正的 DOM 元素。

官方數據要打折扣看

Runway 公布的內部測試顯示，Solaris 在介面還原度（用 SSIM 與 DINOv3 相似度衡量）上贏過 GPT-4o、Gemini 2.5 Pro 與 Fable 5；另外一份 250 人的使用者調查中，61% 的人在「指令跟隨」的比較上更喜歡 Solaris 勝過 Claude Opus 5。這些數字目前都只是 Runway 自己公布的一次性研究，還沒有第三方基準測試驗證，參考就好，別當成定論。

誠實地說：現在能拿來做什麼

還沒開放公測，只能填表申請 early access，一般開發者現在摸不到。
畫面不等於程式碼——因為輸出的是影片幀而非元件，無法直接匯出成可部署的網站或 App，這點跟「AI 生成完整可上線產品」的期待有落差。
比較實際的應用場景是原型展示、電商情境模擬、教學示範，以及訓練其他 AI agent 用的合成介面環境——這些場景本來就不需要「真的能上線的程式碼」。
即時影片生成的算力成本與延遲，目前官方沒有公布，量產部署後的實際體驗跟成本仍是未知數。

對開發者而言，Runway Solaris 代表的是一種新的技術路線，而不是明天就能取代你手上的前端框架。如果你需要的是可維護、可部署的產品，現階段還是老實用 v0 或 Bolt 這類會吐出真程式碼的工具；如果你要做的是快速原型、無互動邏輯要求的展示畫面，或是想餵給 agent 訓練用的合成資料，Solaris 值得申請候補名單先玩玩看。

好不好用，試了才知道。

🇺🇸 Runway Solaris Review: No-Code UI, Frame by Frame

Runway Solaris is the AI tool everyone is talking about this week: announced August 31, 2026, it is billed as the first "Interface World Model" — an AI that draws interactive user interfaces frame by frame, like a video, instead of writing a single line of code. It is a genuinely different approach from every AI website builder that came before it.

What an Interface World Model Actually Does

Tools like v0, Bolt, and Lovable work by having an LLM write out a React or HTML component tree, which your browser then renders. Solaris flips that: it is built on top of Runway's Gen-4.5 video model, and treats your clicks, drags, and typed instructions as inputs for the next video frame. The output is the interface itself — pixels, not markup. The buttons and animations you see are frames the model paints in real time, not real DOM elements.

Take the Benchmarks With a Grain of Salt

Runway's own internal testing claims Solaris beats GPT-4o, Gemini 2.5 Pro, and Fable 5 on interface-reconstruction accuracy (measured via SSIM and DINOv3 similarity). A separate 250-person user study found 61% preferred Solaris's instruction-following over Claude Opus 5. These are Runway's own numbers from a single self-published study — no independent benchmark has verified them yet, so treat the claims as a starting point, not a conclusion.

What You Can Actually Do With It Right Now

It is not publicly available. Access is limited to an early-access waitlist form — most developers cannot touch it yet.
Pixels are not code. Because Solaris outputs video frames rather than components, there is no export path to a deployable website or app. That is a real gap versus the "AI builds a shippable product" pitch.
The realistic use cases right now are prototyping, e-commerce storefront demos, tutorial walkthroughs, and synthetic training environments for other AI agents — situations that do not actually need shippable code.
Runway has not published latency or compute cost for real-time video generation at scale, so the real-world cost of running this is still unknown.

For developers, Solaris represents a new technical direction, not a replacement for your frontend stack tomorrow. If you need something maintainable and deployable, stick with code-generating tools like v0 or Bolt for now. If you want fast prototypes, non-interactive demo screens, or synthetic data to train agents on, Solaris is worth joining the waitlist for.

好不好用，試了才知道 (You will not know if it is good until you try it).

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士