李飞飞最新世界模型正式开放 我用它一句话造出重庆山城Fei-Fei Li's Latest World Model Officially Opens: I Built Chongqing's Mountain City With a Single Sentence


AI free token/AI免费积分

之前凭借「一张图片生成整个 3D 世界」刷屏的李飞飞世界模型,现在终于向所有人开放了。

就在今天,由「AI 教母」李飞飞创办的 World Labs 正式开放世界模型 Atlas,同步上线网页应用和开发者 API,新使用者注册即送 10000 积分。

李飞飞最新世界模型正式开放 我用它一句话造出重庆山城

感受地址:atlas.worldlabs.ai

现在,笔者也能亲自感受这项曾经只能在演示视频里看到的能力:随手画几栋房子、一条小路,AI 就能把草图变成一个可以自由探索的 3D 世界。甚至还能调整情形布局、规划镜头、生成视频,最后导出到 Blender 或游戏引擎里继续加工。

这次开放还带来了两款全新的 Web 应用 Viewfinder 和 Chisel,进入 Atlas 网页直接就能上手感受。

李飞飞转发 World Labs 宣布开放的推文并难掩兴奋,表示「Atlas 现已向所有人开放!!我格外期待看到使用者能用笔者最新的世界模型构建出什么!🤩🌎」

Image

a16z 合伙人 Martin Casado 对这次发布也给出了超高的评价,他认为这是自生成式 AI 兴起以来,3D 生成/重建领域最大的一步。

Image

完整的世界模型工具链

登录 Atlas 需要手机号验证,笔者测试使用 +86 手机号也能正常收到验证码,不过在之后挑选使用区域内,要记得挑选美国或加拿大,以确保后续感受正常。

新使用者登录之后,能看到首页包含四个工具,Viewfinger、Chisel、Relight 和 Splat to mesh,看起来都是很专业的词汇,但实际上就是在达成「世界」从生成的工程化的完整链路。

Image

Viewfinder 的作用是从图片到情形,给它照片或文生成图,它就可以建立情形,让笔者探索新的视角、设置相机路径,再渲染出视频。原本固定在图片里的画面,有了被重新取景的可能。

Image

例如,一张客厅图片就可以成为 Viewfinder 的起点。笔者想看的,可能是从窗边望向沙发的画面,也可能是一段沿着房间缓慢移动的镜头。

但这张照片没有拍到沙发背面,模型就需要推测那里长什么样;画面之外还有多大空间,也需要补全。Viewfinder 是从单张图片扩展,生成的是一个可以延展的情形,包含大量想象。

Image

如果想重建一个现实中存在的地方,就需要给出更多视角。官方也提到,Viewfinder 兼容从单张图片生成情形,也可以使用最多 64 个输入来重建现实环境。

但创作很快会遇到下一个难题:生成的房间确实好看,布局却不符合拍摄需要。

像是沙发离窗户太远,走廊太窄,想要的镜头根本摆不进去。这时候,第二款工具 Chisel 就派上用场了。笔者可以先用简单的三维形状搭出情形结构,确定物体的位置,再让模型依据这些结构生成视觉成效。

笔者可以把它理解成一次简化的布景:先决定墙在哪里、家具怎么放,再考虑材质、光线和总体风格。World Labs 的开发者在演示中,把这套工作方式概括为先搭空间,再自由调整视觉风格,同时保持墙壁的位置。

Image

总的来说,四项工具的能力基本上把世界模型需要用的能力全部达成了。

  • Viewfinder:图片 → 情形。输入 JPG/PNG/WEBP/HEIC,生成新视角(MP4)或高斯溅射点云(PLY/SPZ)。
  • Chisel:智能体驱动的 3D 环境搭建器。兼容图片、文档、网页 URL、网格模型(GLB/glTF/OBJ/FBX)混合输入,用自然语言指挥 AI 搭白模,再由 Atlas 渲染出图、出 splat。
  • Relight:给 splat 情形重新打光,加阴影。
  • Splat to mesh:把 splat 点云转成带贴图的 GLB 网格,直接对接游戏引擎。

Image

当下,整个平台使用积分制:Viewfinder 渲染一帧 50 积分,Chisel 智能体按 token 计费(Gemini 4 Argon,每 1M 输出 token 10000 积分),Chisel 整组渲染 160 积分,生成完整 Splat 则要 5000 积分。

Image

笔者在 Viewfinder 内每生成一张图片是消耗 50 积分,比较大额的积分消耗都在渲染和视频生成上,免费计划领取的 10000 积分,还是可以用来简单感受一下 Atlas 这个新模型的各项能力。

Image

实测 Atlas,重建一座重庆山城

在几个官方演示的案例中,最出圈的一个,大概就是把笔记本上随手涂鸦的简笔画,变成可以走进去的 3D 情形。

Image

笔者也让 AI 生成了一张草图,在这个笔记本涂鸦上,有小房子、大树、微笑的太阳、蜿蜒的小路,然后上传到 Chisel 让它初始化整个情形。

Image

点击「在 Chisel 中构建」后,平台会先生成一张参考图,然后内置智能体开始自动搭建 3D 白模。笔者能实时看到它拆解建筑体量、开门窗、加招牌、布置机位,还会自动检查并修复悬空、穿插等几何错误。

在左侧的智能体对话框内,还能进一步调整 3D 搭建的流程。在搭建完成后,智能体会交付情形、多个零遮挡机位和几套渲染思路,包含不同时间、天气、氛围等。

Image

打开右侧输出面板,挑选一套思路点「渲染」(约 160 积分),即可一次性生成全部机位的成片。

如果还想继续加工,可以给智能体追加指令,例如「加个夜宵小摊」,也可以点「创建视频」生成漫游视频,或生成完整 Splat 后在视口里 360° 漫游。

最后借助 Relight 调整光照氛围,或用 Splat to mesh 转成 GLB 导出。

笔者也尝试了输入一句提示词「暴雨夜的重庆山城巷弄:陡峭的石阶向下延伸,两侧是层层叠叠的老式居民楼,挂满了霓虹灯招牌和晾衣绳,湿漉漉的石板路反射着粉蓝色灯光,远处一列轻轨正从楼体中间穿过。」

大约 1 分钟的时间,Atlas 先生成了一张参考图,里面有石阶、筒子楼、霓虹、晾衣绳、湿路面反光、穿楼轻轨,元素。

Image

Chisel 内置的智能体(Gemini 4 Argon)接管了后续的搭建,它自动制定并执行了 6 步计划,全程约 30 分钟。

涵盖拆解建筑体量 → 布置 9 个机位 → 搭建穿楼轻轨与远景楼群 → 补齐阳台、防盗网、波纹雨棚 → 安放 15 块霓虹灯箱与晾衣绳 → 逐镜头校验视线遮挡等工作。

平台持续自动检查「悬空」「穿插」「镜头遮挡」等几何错误,并在发现「轻轨桥墩悬空 8.25 米」这类难题时当场自行修复。智能体的思考过程,也能展现出这套工具确实很专业,它会说「缩短近景与中景建筑的檐口高度,以露出开阔的轻轨穿楼视野」。

Image

最终情形包含 38 级石阶、左右各 5 栋退台居民楼、3 节带受电弓的轻轨车厢、8 个零遮挡机位。

笔者挑选其中一套「暴雨深蓝夜」渲染思路(160 积分),12 个机位一次出齐:湿滑石阶上粉蓝霓虹的倒影、砖墙上的防盗窗笼、绳上晾晒的 T 恤,全部都有。

Image

最快上手 Atlas 的方法还是使用 Viewfinder,点击「Drop or click」上传一张图片(兼容 JPG/PNG/WEBP/HEIC),挑选 Image 模式导入。

上传的图片会作为一个「画框」浮现在 3D 视口里,点击画布完成放置。接着笔者可以用鼠标移动相机:左键拖拽环绕观察,中键拖拽平移,滚轮缩放。

再把相机移到一个新位置之后,这里注意必须平移离开原始机位,因为深度估计需要视差,原地旋转无法渲染;点击 Render(50 积分),约 1 分钟内生成当前视角的新画面;

Image

接着多换几个角度重复渲染,底部胶片栏会累积新帧;帧数足够后,可以挑选 Splats 模式点「Create splats」生成完整点云(约 15 分钟),或在 Video 模式下生成视频。

最后,也是右上角 Export 可导出 MP4 / PLY / SPZ 文件。

Image

如果不想打开网页,Atlas 还给出了完整的开发者 API,在 Chisel 界面内就有「API 文档」入口,可以把世界生成能力集成进自己的机型管线。

会写代码的 AI,也开始搭情形了

Atlas 开放后,World Labs 的外观师 Ian Curtis 把世界模型、编程模型和网页三维工艺放进了同一个演示里。他用最近热门的 Fable 5.2 + three.js 和 Atlas 结合,搭建了一个第一人称实时漫游的 3D 世界。

李飞飞最新世界模型正式开放 我用它一句话造出重庆山城

这组搭配,也刚好接上了最近 AI 编程热潮里的另一个难题:代码越来越容易生成以后,情形和视觉资产由谁来做?

对于一个交互作品来说,程序运行起来,只完成了其中一部分。使用者最终看到的环境、材质和画面,同样需要制作。

研究者 Ben Mildenhall 在介绍这次发布时,给出了更具体的分工:用 Agent 搭建粗略的三维几何结构,再借助 Atlas 的模型接口控制最终画面的生成。Chisel,以及这次给出的 Codex、Claude Code 集成,都在围绕这个方向外观。

在他看来,完全依靠代码程序化地制作精细画面,成本仍旧很高,两条工艺路线可以互相配合。

编程 Agent 负责搭结构、串联操作,世界模型负责依据空间约束生成视觉内容。原本需要在不同工具之间来回切换的工作,开始有了连接起来的可能。

Atlas 也在尝试把生成结果送进现有的制作流程,官方给出了 Blender 集成,以及将 Gaussian splats 转成网格的工具,方便使用者继续用于游戏、交互环境等项目。

Image

当然,导出一个情形,距离完成一款游戏、一部影片,还有不少工作;像是交互规则、角色行为,以及资产的检查和调整,都需要继续处理。

在世界模型赛道的竞争,也和通用模型一样激烈。Viggle 的 Hang Chu 发布了 JST-2 与 Atlas 的并排比较,强调自家思路直接生成三维内容,并质疑 Atlas 将视图重建为 Gaussian splats 所需的额外步骤和耗时。

李飞飞最新世界模型正式开放 我用它一句话造出重庆山城

Atlas 这次开放的真正意义,可能并不在于单次生成的画质,而在于它把「世界模型」包装成了一条可用的工作流:从一张图、一句话,到可分镜、可渲染、可漫游、可进引擎的 3D 资产,每个环节都有明确的工具、明确的标价。

过去,AI 编程工具可以帮笔者写出一个游戏,却很难独立完成游戏里的情形和视觉资产。现在,编程 Agent 可以负责搭建结构、编写交互,世界模型则负责把这些结构变成可以探索、拍摄和继续加工的三维环境。

对独立开发者来说,这意味着制作游戏、交互作品时,可能不再需要从零手工搭建每一个情形。对于普通人来说,即使不懂三维建模,也终于有机会把脑海里的一个画面,变成可以走进去的世界。

当 AI 已经越来越会写代码,下一步,就是让它把代码里的世界也造出来。

本文来源:https://www.ifanr.com/1689379
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Previously, Li Fei-Fei's world model, which went viral for generating an entire 3D world from a single image, is now finally open to everyone.

Just today, World Labs, founded by "AI godmother" Li Fei-Fei, officially opened its world model Atlas, simultaneously launching web applications and a developer API, with new users receiving 10000 credits upon registration.

Fei-Fei Li's Latest World Model Officially Opens: I Built Chongqing's Mountain City With a Single Sentence

Try it at: atlas.worldlabs.ai

Now, I too can personally experience this capability that was once only seen in demo videos: with just a few houses and a path drawn casually, AI turns the sketch into a freely explorable 3D world. It can even adjust the scene layout, plan camera shots, generate video, and finally export to Blender or a game engine for further processing.

This release also brings two brand-new web applications, Viewfinder and Chisel, which can be tried directly by visiting the Atlas website.

Li Fei-Fei reposted the World Labs tweet announcing the launch and, unable to hide her excitement, said, "Atlas is now open to everyone!! I'm especially excited to see what users will build with my latest world model! 🤩🌎"

Image

a16z partner Martin Casado also gave the launch an extremely high evaluation, saying he believes it is the biggest step in the 3D generation/reconstruction field since the rise of generative AI.

Image

A complete world model toolchain

Logging into Atlas requires phone number verification; I tested with a +86 phone number and received the verification code normally, but when selecting the usage region later, remember to choose the United States or Canada to ensure a smooth subsequent experience.

After logging in, new users will see that the homepage contains four tools: Viewfinger, Chisel, Relight, and Splat to mesh. They may look like very technical terms, but in fact they constitute a complete engineering pipeline that takes a "world" from generation to completion.

Image

Viewfinder's role is to go from image to scene: give it a photo or a text-generated image, and it can build a scene, letting me explore new viewpoints, set camera paths, and render video. The image originally fixed in a picture now has the possibility of being reframed.

Image

For example, a picture of a living room can serve as the starting point for Viewfinder. What I want to see might be the view looking at the sofa from the window, or it could be a slow-moving shot traveling along the room.

But this photo doesn't capture the back of the sofa, so the model needs to infer what that area looks like; the amount of space beyond the frame also needs to be completed. Viewfinder expands from a single image, generating an extendable scene that contains a great deal of imagination.

Image

If you want to reconstruct a real-world location, you need to provide more viewpoints. The company also noted that Viewfinder supports generating a scene from a single image, and can also use up to 64 inputs to reconstruct a real environment.

But creation soon hits the next challenge: the generated rooms do look good, yet their layouts do not meet shooting needs.

For example, the sofa is too far from the window and the hallway is too narrow, so the desired shots simply cannot be composed. At this point, the second tool, Chisel, comes into play. The author can first build the scene structure with simple 3D shapes, determine the positions of objects, and then have the model generate visual results based on these structures.

The author can think of this as a simplified set design: first deciding where the walls go and how the furniture is placed, then considering materials, lighting, and overall style. In the demo, World Labs' developers summarized this workflow as building the space first, then freely adjusting the visual style while keeping the wall positions fixed.

Image

Overall, the capabilities of the four tools essentially fulfill all the functions a world model needs.

Viewfinder: Image → Scene. Input JPG/PNG/WEBP/HEIC, generate new viewpoints (MP4) or Gaussian splatting point clouds (PLY/SPZ).

Chisel: An agent-driven 3D environment builder. It supports mixed inputs of images, documents, web URLs, and mesh models (GLB/glTF/OBJ/FBX), allowing you to direct AI in natural language to build whitebox models, which Atlas then renders into images and splats.

Relight: Re-light splat scenes and add shadows.

Splat to mesh: Convert splat point clouds into textured GLB meshes, ready to connect directly to game engines.

Image

Currently, the entire platform uses a credit-based system: Viewfinder rendering one frame costs 50 credits, the Chisel agent is billed by token (Gemini 4 Argon, 10000 credits per 1M output tokens), a full Chisel group render costs 160 credits, and generating a complete Splat costs 5000 credits.

Image

In Viewfinder, each image the author generates consumes 50 credits, and the larger credit consumption comes from rendering and video generation; the 10000 credits from the free plan are still enough to get a simple feel for the various capabilities of the new Atlas model.

Image

Testing Atlas in practice: Reconstructing a Chongqing mountain city

Among the official demo cases, the one that gained the widest attention is probably the one that turns a casual doodle on a notebook into a 3D scene you can walk into.

Image

The author also had the AI generate a sketch; this notebook doodle featured small houses, big trees, a smiling sun, and a winding path, which was then uploaded to Chisel to initialize the entire scene.

Image

After clicking "Build in Chisel," the platform first generates a reference image, and then the built-in agent automatically begins constructing the 3D whitebox. The author can watch in real time as it breaks down building volumes, opens doors and windows, adds signboards, and sets up camera positions, while automatically checking and fixing geometric errors such as floating and intersecting elements.

In the agent dialog box on the left, you can further adjust the 3D construction workflow. Once the construction is complete, the agent will deliver the scene, multiple unobstructed camera positions, and several rendering concepts covering different times, weather, atmospheres, and more.

Image

Open the output panel on the right, select a concept and click "Render" (approximately 160 credits) to generate the final footage for all camera positions in one go.

If you want to keep refining, you can give the agent additional instructions, such as "add a late-night food stall." You can also click "Create Video" to generate a roaming video, or generate a complete Splat and take a 360° tour in the viewport.

Finally, use Relight to adjust the lighting atmosphere, or use Splat to mesh to convert it into a GLB file for export.

The author also tried inputting the prompt: "A rainy night in the alleys of Chongqing's mountain city: steep stone steps extend downward, with layers of old residential buildings on both sides, hung with neon signs and clotheslines, the wet stone paths reflecting pink and blue lights, and a light rail train passing through the middle of the buildings in the distance."

In about 1 minute, Atlas produced a reference image containing the elements of stone steps, tube buildings, neon lights, clotheslines, reflections on wet roads, and a light rail passing through buildings.

Image

The built-in agent in Chisel (Gemini 4 Argon) took over the subsequent construction, automatically formulating and executing a 6-step plan that took approximately 30 minutes in total.

The tasks included breaking down the building masses → arranging 9 camera positions → constructing the light rail passing through the buildings and the distant building clusters → adding balconies, security grilles, and corrugated rain canopies → placing 15 neon light boxes and clotheslines → and checking the line of sight and obstructions for each shot.

The platform continuously and automatically detects geometric errors such as "floating," "interpenetration," and "camera occlusion," and automatically repairs them on the spot upon discovering complex issues like "light rail bridge piers floating 8.25 meters." The AI agent's thought process also demonstrates the true professionalism of this tool, as it will state, "Reduce the eave heights of foreground and midground buildings to reveal an expansive view of the light rail passing through the buildings."

Image

The final scene includes 38 stone steps, 5 terraced residential buildings on each side, 3 light rail cars with pantographs, and 8 unobstructed camera positions.

The author selected one of the "heavy rain, deep blue night" rendering concepts (160 credits), and the 12 camera positions were generated at once: the pink and blue neon reflections on the slippery stone steps, the security window grilles on the brick walls, and the T-shirts drying on the ropes were all included.

Image

The fastest way to get started with Atlas is to use Viewfinder; click "Drop or click" to upload an image (compatible with JPG/PNG/WEBP/HEIC), select Image mode, and import it.

The uploaded image will appear as a "frame" floating in the 3D viewport; click the canvas to place it. Then the author can move the camera with the mouse: left-click and drag to orbit, middle-click and drag to pan, and use the scroll wheel to zoom.

Next, move the camera to a new position. Here it is important to pan away from the original camera position, because depth estimation requires parallax, and rotating in place cannot render; click Render (50 credits) to generate a new image of the current view within about 1 minute;

Image

Then, render repeatedly from several other angles, and the film strip at the bottom will accumulate new frames; once there are enough frames, you can select Splats mode and click "Create splats" to generate a complete point cloud (approximately 15 minutes), or generate a video in Video mode.

Finally, the Export button in the upper right corner allows you to export MP4 / PLY / SPZ files.

Image

If you don't want to open a webpage, Atlas also provides a complete developer API, with an "API Documentation" entry available within the Chisel interface, allowing you to integrate world generation capabilities into your own model pipeline.

AI that can write code has also begun building scenes.

After Atlas was opened up, Ian Curtis, a visual designer at World Labs, put world models, programming models, and web 3D craft into a single demo. He combined the recently popular Fable 5.2 + three.js with Atlas to build a 3D world with first-person real-time roaming.

Fei-Fei Li's Latest World Model Officially Opens: I Built Chongqing's Mountain City With a Single Sentence

This combination also happens to address another challenge in the recent AI programming boom: once code becomes increasingly easy to generate, who will handle the scenes and visual assets?

For an interactive work, having the program run is only part of the completion. The environment, materials, and visuals that users ultimately see also need to be produced.

When introducing this release, researcher Ben Mildenhall outlined a more specific division of labor: using Agents to build rough 3D geometric structures, then leveraging Atlas's model interface to control the generation of the final visuals. Chisel, along with the Codex and Claude Code integrations introduced this time, are all built around this direction.

In his view, relying entirely on code to procedurally produce fine-grained visuals still carries high costs, and the two craft routes can complement each other.

Programming Agents are responsible for building structures and chaining operations, while world models are responsible for generating visual content based on spatial constraints. Work that previously required switching back and forth between different tools is now beginning to have the possibility of being connected.

Atlas is also trying to feed generation results into existing production pipelines, with the official release offering Blender integration as well as a tool to convert Gaussian splats into meshes, making it convenient for users to continue using them in projects such as games and interactive environments.

Image

Of course, exporting a scene is still far from completing a game or a film; there is still quite a bit of work to do, such as interaction rules, character behavior, and the inspection and adjustment of assets, all of which need to be further processed.

Competition in the world model track is just as intense as in general models. Hang Chu of Viggle released a side-by-side comparison of JST-2 and Atlas, emphasizing that his own approach directly generates 3D content, and questioning the additional steps and time required for Atlas to reconstruct views as Gaussian splats.

Fei-Fei Li's Latest World Model Officially Opens: I Built Chongqing's Mountain City With a Single Sentence

The true significance of Atlas's open release this time may not lie in the quality of a single generation, but in the fact that it packages the "world model" into a usable workflow: from a single image or a sentence to 3D assets that can be storyboarded, rendered, roamed through, and imported into engines, with clear tools and clear pricing at every stage.

In the past, AI programming tools could help the author write a game, but were hard-pressed to independently complete the scenes and visual assets within it. Now, programming Agents can be responsible for building structures and writing interactions, while world models are responsible for turning those structures into three-dimensional environments that can be explored, filmed, and further processed.

For independent developers, this means that when creating games and interactive works, they may no longer need to manually build every scene from scratch. For ordinary people, even those who don't know 3D modeling, there is finally an opportunity to turn an image in their mind into a world they can walk into.

As AI becomes increasingly capable of writing code, the next step is to let it also build the world within the code.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry