The World Labs Unveils Atlas World Model — A Deep Dive into the Four-Modal Spatial Foundation Model
The World Labs Unveils Atlas World Model — A Deep Dive into the Four-Modal Spatial Foundation Model
Abstract: On September 1, 2026, World Labs, co-founded by Fei-Fei Li, officially launched Atlas — a multimodal autoregressive diffusion transformer pretrained from scratch, natively supporting four modalities: text, images, video, and 3D. Atlas demonstrates capabilities surpassing specialized models in pixel-perfect camera control, sparse-view 3D reconstruction, and Real-to-Sim robotics simulation, marking a pivotal step in spatial intelligence moving from academic concept to engineering infrastructure.
1. Background: Spatial Intelligence from Vision to Product
1.1 The ImageNet Legacy and the Ambition of Spatial Intelligence
Fei-Fei Li’s track record in computer vision is nearly unparalleled. She created the ImageNet dataset that catalyzed the deep learning revolution, served as director of the Stanford AI Lab (2013-2018), was VP and Chief Scientist of AI/ML at Google Cloud, and is a founding co-director of Stanford’s Institute for Human-Centered Artificial Intelligence (HAI). The journey from ImageNet to World Labs follows a clear technical progression: first teaching machines to recognize what appears in an image, then building systems that can represent the three-dimensional space beyond visible pixels.
Founded in 2024 in San Francisco, World Labs was co-founded by Fei-Fei Li alongside Justin Johnson (Stanford Ph.D., working on visual reasoning, image generation, and 3D reasoning), Ben Mildenhall (co-creator of NeRF, former Google research scientist), and Christoph Lassner (former research lead at Meta Reality Labs and Epic Games). This team composition represents a golden combination — the brightest minds in visual understanding, 3D reconstruction, and real-time rendering gathered under one roof.
1.2 The Evolution from Marble to Atlas
Atlas did not emerge from a vacuum. It builds upon Marble, World Labs’ first commercial multimodal model:
- November 2025: Marble officially launched, supporting 3D world creation from multimodal inputs, validating the spatial intelligence framework
- January 2026: World API launched, providing developers with programmatic access to spatial intelligence tools
- April 2026: Marble 1.1 and Marble 1.1-Plus released, improving image quality and large-scale scene generation
- July 2026: World Labs acquired SceniX, strengthening robotics-focused spatial intelligence
- September 1, 2026: Atlas officially unveiled
1.3 Funding Scale Reveals Industry Confidence
World Labs has raised a total of $1.23 billion:
| Round | Amount | Key Investors |
|---|---|---|
| 2024 Seed | $230M | Andreessen Horowitz, NEA, Radical Ventures |
| February 2026 | $1B | AMD, Autodesk ($200M), NVIDIA, Fidelity, Emerson Collective, Sea |
(Source: World Labs official announcement, https://www.worldlabs.ai/blog/atlas)
This funding scale is exceptionally rare for an AI startup. Autodesk’s $200M contribution is particularly noteworthy — it signals that Atlas’s technical roadmap is deeply tied to the digital content creation (DCC) tool ecosystem.
2. Architecture Deep Dive: Four-Modal Autoregressive Diffusion Transformer
2.1 Architecture Overview
Atlas’s core architecture can be summarized as: Multimodal + Autoregressive + Diffusion + Transformer. These four technical pillars work together to form a unified Spatial Context paradigm.
+---------------------------------------------------------------------------+
| Atlas Architecture Overview |
+---------------------------------------------------------------------------+
| |
| +----------+ +---------------+ +------------------+ |
| | Text | | | | Autoregressive | |
| | (Prompt) |--->| | | (Token-by-Token) | |
| +----------+ | | +------------------+ |
| | Images | | Spatial |--->| Rectified Flow | |
| | (1-6) |--->| Context | | Diffusion | |
| +----------+ | | +------------------+ |
| | Camera |--->| | | Transformer | |
| | Poses | | | | Backbone | |
| +----------+ | | +------------------+ |
| | Depth |--->| | | |
| | Maps | | | v |
| +----------+ +---------------+ +------------------+ |
| | Outputs | |
| | +- Images/Video | |
| | +- 3D Point Cloud| |
| | +- 3D Gaussian | |
| | Splats | |
| +------------------+ |
+---------------------------------------------------------------------------+
2.2 Spatial Anchoring of Multimodal Inputs
Unlike traditional LLMs that encode all inputs into a unified token sequence, Atlas’s core innovation is that each image is anchored at a specific position in 3D space. This means:
- Inputs include text, images, camera poses, and 3D depth maps
- Videos are represented as sequences of images, each with a corresponding camera pose
- Each image and depth map is explicitly conditioned on its camera pose
+---------------------------------------------------------------------------+
| Spatial Context Construction |
+---------------------------------------------------------------------------+
| |
| +-----+ +-----+ +-----+ +-----+ |
| |Img1 | |Img2 | |Img3 | |Img4 | ... |
| |Pose1| |Pose2| |Pose3| |Pose4| |
| +--+---+ +--+---+ +--+---+ +--+---+ |
| | | | | |
| v v v v |
| +-------------------------------------------------------------------+ |
| | 3D Spatial Position Encoding Layer | |
| +-----------------------------------+-----------------------------------+ |
| | |
| v |
| +-------------------------------------------------------------------+ |
| | Unified Spatial Context Representation | |
| +-------------------------------------------------------------------+ |
+---------------------------------------------------------------------------+
2.3 Autoregressive Generation: Sequential World Construction
Atlas transforms world modeling into a multimodal sequence generation problem. Each element is one of the four modalities, outputs are generated one at a time, conditioned on the preceding sequence.
This design naturally adapts to diverse tasks — each task is simply a different type of sequence, with inputs followed by outputs.
# Atlas autoregressive generation workflow
class AtlasAutoregressiveGeneration:
def __init__(self, model, tokenizer):
self.model = model
self.tokenizer = tokenizer
def generate_sequence(self, task_type, inputs):
sequence = []
for img, pose in zip(inputs['images'], inputs['camera_poses']):
tokens = self.tokenizer.encode_image(img, pose)
sequence.append(tokens)
spatial_context = self.model.encode_spatial_context(sequence)
outputs = []
for step in range(self.max_steps):
next_element = self.model.diffuse_step(
context=spatial_context, prev_outputs=outputs,
cfg_scale=7.5, num_steps=50)
outputs.append(next_element)
spatial_context = self.model.update_kv_cache(
spatial_context, next_element)
return outputs
2.4 Diffusion Generation: Rectified Flow Model
Atlas uses a Rectified Flow model, generating outputs by gradually denoising them. Diffusion models excel at modeling high-dimensional continuous data like images and video, and can naturally trade off speed and quality by varying the number of denoising steps during inference.
Key characteristics:
- Latent Diffusion: Operates in VAE-encoded latent space, not raw pixel space
- Classifier-Free Guidance (CFG): Controls the balance between generation and condition adherence
- Shifted Noise Schedules: Optimized noise schedules for different data types
+---------------------------------------------------------------------------+
| Rectified Flow Diffusion Process |
+---------------------------------------------------------------------------+
| |
| Pure Noise step 10 step 30 step 50 |
| +-----+ +-----+ +-----+ +-----+ |
| | ## | --> | .. | --> | .. | --> | ## | |
| | #### | | .... | | .... | | .... | |
| |######| |......| | .... | | .... | |
| +-----+ +-----+ +-----+ +-----+ |
| ^ ^ ^ ^ |
| | | | | |
| z_T ~ N(0,I) Rectified Flow z_{t} z_0 |
| ODE |
| |
| d z_t = v_theta(z_t, t, context) dt |
+---------------------------------------------------------------------------+
2.5 Transformer Backbone: Merging LLM and Video Model Advantages
Atlas’s Transformer architecture benefits from advances in both LLMs and video models:
From LLMs:
- KV-caching for accelerated inference
- Cache-aware routing
- Disaggregated serving
From Video Models:
- Diffusion distillation
- Classifier-free guidance
- Shifted noise schedules
- VAE design improvements
import torch
import torch.nn as nn
class AtlasTransformerBlock(nn.Module):
def __init__(self, d_model=4096, n_heads=32, d_ff=16384):
super().__init__()
self.spatial_attention = nn.MultiheadAttention(
d_model, n_heads, batch_first=True)
self.spatial_pos_encoding = nn.Sequential(
nn.Linear(7, d_model), nn.SiLU(), nn.Linear(d_model, d_model))
self.ffn = nn.Sequential(
nn.Linear(d_model, d_ff), nn.GELU(), nn.Linear(d_ff, d_model))
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
def forward(self, x, camera_poses):
spatial_pos = self.spatial_pos_encoding(camera_poses)
x = x + spatial_pos
attn_out, _ = self.spatial_attention(
self.norm1(x), self.norm1(x), self.norm1(x))
x = x + attn_out
ffn_out = self.ffn(self.norm2(x))
x = x + ffn_out
return x
3. Capability Overview
3.1 Pixel-Perfect Camera-Controlled Generation
This is Atlas’s most striking capability. Traditional video generation models convey camera motion intent through text descriptions, while Atlas directly receives precise camera geometry as a native input type.
Specifications:
- Input: 1-6 reference images + manually designed camera path
- Output: Up to 1440p resolution, up to 1 minute of video
- Scene types: Indoor/outdoor, realistic/stylized, static/dynamic
+---------------------------------------------------------------------------+
| Camera-Controlled Generation vs Text Prompt |
+---------------------------------------------------------------------------+
| |
| Traditional: "Camera pans left 30 degrees" (text description) |
| | |
| v |
| [Video Model] --> Fuzzy camera control, multiple attempts needed |
| |
| Atlas: [Camera Pose Matrix] |
| [[1.0, 0.0, 0.0, 2.5], <- Position + Rotation Matrix |
| [0.0, 0.87, -0.5, 1.2], |
| [0.0, 0.5, 0.87, 0.0]] |
| | |
| v |
| [Atlas] --> Pixel-perfect camera control, one-shot success |
+---------------------------------------------------------------------------+
Spatial Anchoring: A particularly impressive capability is that you can place two unrelated reference images at different positions in 3D space, and Atlas will generate a smoothly interpolated world between them — it imagines doorways, hallways, nooks, and other transitional structures.
3.2 Sparse-View 3D Reconstruction
Atlas can reconstruct real-world scenes from one to dozens of input images. Crucially, it requires no special capture equipment or hundreds of dense views.
Reconstruction capabilities:
- 2-3 images are typically sufficient for faithful reconstruction
- More input images reduce the parts the model needs to imagine
- Output formats: Point clouds or 3D Gaussian Splats
+---------------------------------------------------------------------------+
| Sparse-View 3D Reconstruction Pipeline |
+---------------------------------------------------------------------------+
| |
| Input Images (2-3) Output 3D Point Cloud |
| +------+ +----------------------------------+ |
| | PH | --> Feature Extract | o o o o o o o o o | |
| | PH | --> Depth Estimate | o o o o o o o o o | |
| | PH | --> Point Cloud | o o o o o o o o o | |
| +------+ | o o o o o o o o o | |
| | +----------------------------------+ |
| v | |
| +----------------------+ v |
| | Atlas World Knowledge | +----------------------------------+ |
| | (Fills unseen areas) | | 3D Gaussian Splats | |
| +----------------------+ | (Interactive Rendering) | |
| +----------------------------------+ |
+---------------------------------------------------------------------------+
Impact of input image count on reconstruction quality:
def analyze_reconstruction_quality(atlas_model, scene_images):
"""Analyze how input image count affects reconstruction quality"""
results = {}
for n in [1, 2, 3, 5, 10, 25]:
subset = scene_images[:n]
point_cloud = atlas_model.reconstruct_3d(
images=subset, cameras=[img.camera for img in subset])
results[n] = {
'f_score': point_cloud.eval_f_score(),
'hallucination_ratio': point_cloud.eval_hallucination()}
print(f'Views: {n:2d} | F-Score: {results[n]["f_score"]:.3f}')
return results
3.3 Space-Time Simulation
Atlas simultaneously understands the spatial structure of the world and how it evolves over time, enabling two important applications:
Bullet Time Reframing: With just 3-5 ordinary cell phone cameras, Atlas can freeze time and reframe shots, letting you observe events from impossible angles. No professional photography equipment is needed — each clip was captured by engineers using tripods and clamps that fit in a backpack.
Robotics Simulation (Real-to-Sim): After reconstructing an environment from phone video, Atlas generates the RGB images and depth readings that a robot-mounted camera would observe while moving through the space. For manipulation tasks, it can reconstruct the visual and geometric properties of rigid, articulated, and deformable objects.
+---------------------------------------------------------------------------+
| Real-to-Sim Robotics Pipeline |
+---------------------------------------------------------------------------+
| |
| Real-World Recording Simulation Environment |
| +--------------+ +----------------------------------+ |
| | Phone Video | --> 3D | Atlas 3D Scene Reconstruction | |
| | (24 frames) | Rec. | (Point Cloud / Gaussian Splats) | |
| +--------------+ +----------------------------------+ |
| | |
| v |
| +-----------------------+ |
| | Robot Trajectory Plan | |
| +-----------------------+ |
| | |
| v |
| +-------------------------------------------------------------------+ |
| | Atlas Simulated Sensor Output: RGB + Depth (along robot path) | |
| +-------------------------------------------------------------------+ |
| | |
| v |
| +-------------------------------------------------------------------+ |
| | Controllable Variations: Object/Lighting/Background/Robot/Physics | |
| +-------------------------------------------------------------------+ |
| | |
| v |
| +-------------------------------------------------------------------+ |
| | Output: Diverse Training Data + Test Environments (Robot Learning) | |
| +-------------------------------------------------------------------+ |
+---------------------------------------------------------------------------+
3.4 Image Generation and 360-degree Panoramas
While world modeling is Atlas’s primary focus, it is also a capable image generator: it follows complex prompts, renders text, and generates a wide variety of visual styles, including 360-degree panoramas from text or image prompts.
4. Benchmark Analysis: Strengths and Limitations
4.1 Camera-Controlled Generation Evaluation
World Labs conducted an internal evaluation comparing Atlas against five top-tier video models. Methodology: single input image + 1-3 cinematic camera motions, Atlas using native camera input format, other models using text descriptions.
Third-party human raters judged which model better followed the intended camera path:
| Comparison Model | Raters Preferring Atlas |
|---|---|
| MiniMax H3 | 75% |
| Gemini Omni Flash | 81% |
| Happy Horse 1.1 | 86% |
| FLUX 3 | 93% |
| Seedance 2.5 | 94% |
(Source: World Labs official blog, https://www.worldlabs.ai/blog/atlas)
Key interpretation: Atlas’s advantage grows with more complex camera trajectories. However, competing models received camera instructions through text descriptions, which inherently loses information. World Labs acknowledges that more sophisticated prompt engineering could improve other models’ camera following.
4.2 3D Reconstruction Evaluation
On sparse-view 3D reconstruction, Atlas was compared against five specialized open-source reconstruction models:
+---------------------------------------------------------------------------+
| 3D Reconstruction Error (AbsRel x 10^-3, lower is better) |
+---------------------------------------------------------------------------+
| |
| Atlas (Ours) ############################ 25.3 |
| Pi3X (posed) ################################ 28.7 |
| pi^3 ################################## 30.1 |
| VGGT-Omega 1B ######################################## 37.2 |
| Depth Anything 3 ############################################ 42.5 |
| MapAnything ################################################## 47.7|
| |
| Datasets: DTU / ETH3D / KITTI / NRGBD / 7-Scenes / T&T / ScanNet |
+---------------------------------------------------------------------------+
Notable detail: On one test set, the open-source model VGGT-Omega 1B achieved lower error (40.2) compared to Atlas (42.4). World Labs emphasizes they reproduced all baseline results to ensure fair comparison.
(Source: World Labs official blog, https://www.worldlabs.ai/blog/atlas)
4.3 Caveats to Consider
- Internal evaluation, not independent third-party validation: All evaluations conducted by World Labs internally, no independent replication yet
- Inherent asymmetry in camera test: Atlas receives native camera format, competitors receive only text descriptions
- Early access limitations: Currently limited to select partners, no public pricing, no GA date
- Self-reported scaling laws: The company claims larger models perform better, but this is their own result, not independent evidence
5. Commercial Prospects and Competitive Landscape
5.1 Productization Path: From Atlas to Marble
Atlas will power future versions of Marble, World Labs’ 3D persistent environment generation product. Marble already has public pricing:
| Tier | Monthly Price |
|---|---|
| Standard | $20 |
| Pro | $35 |
| Max | $95 |
| Enterprise | Custom pricing |
(Source: RuntimeWire, https://runtimewire.com/article/world-labs-atlas-spatial-intelligence-world-model)
5.2 Four Routes in the World Model Landscape
The current world model field comprises four distinct technical approaches:
| Approach | Representative | Core Philosophy |
|---|---|---|
| Pixel Generation | Seedance, Sora | Simulate world through video generation, emphasizing visual quality |
| Spatial Intelligence | World Labs/Atlas | Start from 3D structure, emphasizing geometric precision |
| Physics Simulation | Odyssey, LeCun’s AMI Labs | Build interactive world simulations, emphasizing physical realism |
| Geospatial | Niantic Spatial | Focus on geospatial mapping and visual positioning |
(Source: 36Kr, https://36kr.com/p/3963266346014344)
Atlas’s core differentiation: A single model unifying spatially-anchored multimodal inputs, producing video, explicit 3D geometry, and simulated sensor views.
5.3 Potential Challenges
- Boundary between geometric consistency and hallucination: Can Atlas maintain geometric accuracy when reference images disagree? Needs partner workload validation
- Sim-to-Real gap: Are generated sensor views accurate enough for real robot hardware?
- Data flywheel: As a closed-source model, how does Atlas collect user feedback and continuously improve?
- Competitive acceleration: Odyssey, LeCun’s AMI Labs, and others are iterating rapidly, with some choosing open-source routes
6. Technical Summary and Outlook
6.1 Why Atlas Matters
Atlas represents a paradigm shift from AI generating pixels to AI understanding space. It is not a video generation model, nor a 3D reconstruction model — it is a spatial foundation model that treats camera geometry as a first-class citizen and 3D consistency as a core constraint rather than a post-processing step.
6.2 Technology Roadmap Outlook
+---------------------------------------------------------------------------+
| Atlas Technology Roadmap (Speculative) |
+---------------------------------------------------------------------------+
| |
| Phase 1 (Current): Four-Modal Foundation Model |
| +- Camera-Controlled Generation (1440p, 1min) |
| +- Sparse-View 3D Reconstruction |
| +- Space-Time Simulation (Bullet Time) |
| +- Image / Panorama Generation |
| |
| Phase 2 (Near-term): Physical Interaction Enhancement |
| +- Rigid/Articulated/Deformable Object Physics Simulation |
| +- Longer Video Generation (>1min) |
| +- Real-time Interactive World Exploration |
| |
| Phase 3 (Mid-term): Embodied Intelligence Integration |
| +- End-to-End Robot Policy Learning |
| +- Multi-Agent Simulation Environments |
| +- Closed-Loop Real2Sim2Real |
| |
| Phase 4 (Long-term): General Spatial Reasoning |
| +- Causal Reasoning and Physical Commonsense |
| +- Open-World Understanding and Planning |
| +- Cross-Scene Knowledge Transfer |
+---------------------------------------------------------------------------+
6.3 Impact on the AI Industry
Atlas’s launch is not just a milestone for World Labs — it may have far-reaching implications for the entire AI industry:
- Film and VFX: Directors and VFX artists gain precise virtual camera control without complex 3D scene reconstruction pipelines
- Gaming: The pipeline from concept art to explorable 3D environments is dramatically shortened
- Robotics: A potential solution to the data scarcity problem — from phone video to simulation training data
- Architecture/Real Estate: Generating complete 3D building models from a handful of photos becomes feasible
7. Code Example: Using the Atlas API (Simulated)
The following code simulates Atlas API calls, demonstrating how to use it in a real workflow:
import numpy as np
from typing import List, Optional
from dataclasses import dataclass
@dataclass
class CameraPose:
position: np.ndarray
rotation: np.ndarray
timestamp: float = 0.0
@dataclass
class AtlasConfig:
model_version: str = "atlas-v1"
resolution: str = "1440p"
num_diffusion_steps: int = 50
cfg_scale: float = 7.5
max_video_duration: int = 60
class AtlasClient:
def __init__(self, api_key: str, config: Optional[AtlasConfig] = None):
self.api_key = api_key
self.config = config or AtlasConfig()
def camera_controlled_generation(self, reference_images, camera_path):
return {"status": "success", "task": "camera_control"}
def sparse_view_reconstruction(self, images, camera_poses):
return {"status": "success", "output_type": "gaussian_splat"}
def real_to_sim(self, phone_video_frames, robot_trajectory):
return {"status": "success", "num_environments": 10}
def demo_camera_control():
client = AtlasClient(api_key="your-api-key")
ref_img = np.random.randn(3, 720, 1280)
camera_path = [CameraPose(np.array([5*np.cos(t), 2, 5*np.sin(t)]),
np.array([0, t, 0]))
for t in np.linspace(0, 2*np.pi, 300)]
result = client.camera_controlled_generation([ref_img], camera_path)
print(result)
8. Key Insights and Industry Impact
8.1 Fei-Fei Li’s Spatial Intelligence Narrative
Fei-Fei Li categorizes world models by function into three types:
- Renderer: Generates visual outputs
- Simulator: Outputs precise physical data
- Planner: Directly guides actions
She believes the simulator is the most critical component for advancing physical AI. Atlas is the engineering realization of this vision — it does not just generate beautiful images, but constructs measurable, geometrically consistent spatial representations.
(Source: 36Kr, https://36kr.com/p/3963266346014344)
8.2 Core Differentiation
Unlike the pixel fitting approach of Sora and similar models, Atlas starts from 3D geometry:
- Sora’s philosophy: With enough pixels, 3D worlds will emerge
- Atlas’s philosophy: Explicit 3D geometry is a core constraint on inputs and outputs, not an emergent byproduct
8.3 Open Challenges
Despite its technical impressiveness, Atlas faces several key challenges:
- Closed-source strategy: Not releasing model weights limits academic research and validation
- Early access: Number of partners undisclosed, independent validation has not occurred
- Pricing uncertainty: Atlas itself has no public price or GA date
- Depth of physics simulation: 3D reconstruction does not equal physical understanding
Conclusion
Atlas is a significant milestone in the field of spatial intelligence. It demonstrates that anchoring multimodal inputs through explicit 3D geometry in a unified spatial context can achieve generation and reconstruction capabilities that surpass specialized models. But there is still a long road from technical demonstration to productization — the workloads of early partners will determine whether Atlas evolves from an impressive demo into indispensable infrastructure.
For AI practitioners, whether you work in film production, game development, robotics, or architectural visualization, Atlas deserves close attention. Spatial intelligence may become the next wave of AI after language intelligence, and World Labs has already secured a first-mover position.
References:
- World Labs, Atlas: A World Model for Spatial Intelligence, https://www.worldlabs.ai/blog/atlas
- RuntimeWire, World Labs launches Atlas for video, 3D reconstruction and robot simulation, https://runtimewire.com/article/world-labs-atlas-spatial-intelligence-world-model
- World Labs, Building Worlds That Train Robots, https://www.worldlabs.ai/blog/real-to-sim-to-real
- World Labs, World Labs Announces New Funding, https://www.worldlabs.ai
- 36Kr, The awkward world model: four routes unresolved, https://36kr.com/p/3963266346014344
- World NL, World Labs unveils Atlas, https://worldnl.com/world-labs-unveils-atlas