Host: In the journey toward spatial intelligence, creating pixels that are spatially contextualized and reasoned is a fundamentally difficult step, and that is what Atlas has achieved. As we know, Large Language Models (LLMs) are built on predicting the next token; video models are built on predicting the next frame. Atlas is fundamentally about predicting novel views. This is a real domain where AI can bring massive benefits to people and production pipelines, slashing costs by 50x or 100x. In the first movie of The Matrix, there is that famous bullet-time shot where Neo leans back. Back then, they required hundreds of physical cameras, green screens, and extremely expensive calibrations to pull that off.

Justin Johnson: Exactly. And with Atlas, we can achieve that with just three ordinary devices—say, three iPhones on tripods. No dedicated studio shooting, no green screens, and no expensive multi-camera calibration required.

Ben Mildenhall: Yes, imagine capturing dynamic moments like someone throwing a ball into a hoop or a strawberry dropping into a bowl of milk with splashes. From just three iPhone video feeds, we can reframe and reconstruct the entire scene from any novel viewpoint. You can freeze time while having a virtual camera fly through the airborne milk splashes, creating that breathtaking frozen-time cinematic effect. In fact, in certain setups, we can even do it with just two cameras.

Host: Could you explain in the simplest, most intuitive terms what Atlas fundamentally does? What are the inputs and outputs of the model?

Justin Johnson: One of the fundamental principles of Atlas is the prediction of novel views. This is the core capability that we found completely fascinating and fundamentally new—no prior model operated this way. LLMs predict the next token, video models predict the next frame, whereas Atlas takes a set of views or descriptions of a scene and maps them into what we call a spatial context. This spatial context implicitly and faithfully represents the 3D physical world. You can then place a virtual camera at any arbitrary point in space and time, and Atlas understands and renders exactly what the world should look like from that precise pose.

Host: Ben, many video models nowadays claim the title of "world models" and claim to generate new camera angles. Could you explain more specifically how Atlas fundamentally differs from previous generative video models?

Ben Mildenhall: The distinction lies in the spatial context Justin mentioned. Many video models gained popularity through image-to-video generation or interpolating between first and last frames. Some newer models accept 20, 30, or 50 reference images. But the defining trait of Atlas is that every input frame is grounded in a rigorous spatial distribution with 3D camera poses.

In conventional models, images are treated as loose visual cues that the model interprets freely based on text prompts. In Atlas, every image comes with precise camera positioning in 3D space. This enables high-accuracy metric reconstruction. If you take four photos from the four corners of a room and feed them into Atlas, it reconstructs an exact 3D replica of what exists, without hallucinating inconsistent objects or broken spatial relationships in unseen corners. It faithfully adheres to spatial reality.

Justin Johnson: At the same time, this unlocks massive creative potential. If you feed in just two sparse, wide-baseline photos, classical 3D reconstruction pipelines (like traditional SfM or standard NeRF) would completely break down due to insufficient overlap. Atlas, as a generative spatial foundation model, uses its rich learned prior to hallucinate plausible structure in unobserved areas while remaining strictly anchored to the geometric constraints of the observed views.

Host: That is truly remarkable. What is the limit of Atlas's performance under extremely sparse views? When given only a handful of photos, how does it maintain photorealism while preserving global geometric consistency?

Ben Mildenhall: Classical 3D reconstruction requires dozens or hundreds of densely captured photos with over 80% overlap, followed by heavy offline bundle adjustment. When views are sparse, traditional geometry-based methods fail catastrophically, producing floating floaters, severe artifacts, and broken surfaces.

Atlas has learned the structural and physical priors of the 3D world during large-scale pretraining. When given only one, two, or three sparse shots, it infers plausible back-surface geometry, consistent lighting, and occlusion relationships. Even if you orbit the camera 180 degrees behind an object, Atlas synthesizes a coherent, continuous, and visually compelling novel view.

Justin Johnson: We observed a very clear scaling law throughout our training runs: as model capacity grows and the scale of high-quality multi-view spatial training data expands, novel view synthesis capabilities improve monotonically. The model preserves fine details—such as typography, specular reflections, and thin structures—while developing an increasingly robust grasp of macro-scale 3D geometry.

Host: Does this mean we will soon experience true 4D video and fully interactive worlds through Atlas? Will users be able to freely walk around dynamic scenes?

Ben Mildenhall: 4D spatiotemporal modeling is a core milestone on our roadmap. In the real world, everything evolves—people move, water splashes, lighting changes. Static 3D only freezes a single instant, whereas 4D spatial intelligence requires joint reasoning across both the temporal axis and spatial camera trajectories.

When training Atlas on real-world datasets, the model learns to decouple dynamic foreground motions from static background geometry. This allows it to either isolate clean static environments or synthesize full 4D spatiotemporal trajectories. While our current release focuses heavily on metric static reconstruction and high-fidelity camera control, the underlying model already encodes profound 4D dynamic representations.

Justin Johnson: I want to emphasize that bridging academic research and industrial-grade deployment hinges entirely on fine-grained editability and controllability. In academia, many demos are essentially toys: you type a prompt, a mesh appears, but you have zero control over metric dimensions, camera extrinsics, or temporal synchronization.

In professional VFX and production pipelines, control is paramount. Creators need to specify exact 3D coordinates, exact asset placement, and exact focal lengths and camera trajectories without degrading rendering fidelity. Maintaining an uncompromisingly high quality bar while introducing multi-modal spatial control is the central engineering moat of Atlas.

Host: This touches upon a deep philosophical question. Beyond faster inference and higher resolution, what broader vision do you see regarding the fundamental nature of spatial intelligence?

Justin Johnson: This brings us back to the fundamental nature of intelligence. True general intelligence cannot merely sit behind a screen parsing tokens; it must interact with the 3D physical world and close the perceptual-action loop.

In computer science, there is the concept of AI-completeness, analogous to NP-completeness in complexity theory where solving 3-SAT allows you to solve any NP problem via reduction. In NLP, predicting the next token is widely viewed as an AI-complete primitive: if a model can flawlessly predict the next token at the end of a mystery novel ("The killer was...") or predict the next mathematical symbol proving the Riemann Hypothesis, it has effectively mastered general intelligence.

Ben Mildenhall: We firmly believe that generative novel view prediction is equally an AI-complete task.

Look at it through the lens of evolutionary biology: why did nature bestow eyes upon animals that move, but not upon trees rooted in the ground? Because locomotion fundamentally necessitates the continuous processing of novel views. To navigate and survive, an organism must maintain a rich internal 3D model of reality, accurately anticipating occlusions, geometry, and causal physical dynamics from unseen perspectives.

If a generative model can flawlessly predict any novel view of the physical world, it inherently implies that the model has internalized 3D spatial geometry, optics, temporal coherence, and physical causality. Therefore, predicting the next view is fundamentally just as profound as predicting the next token.

Host: That is an extraordinarily profound and visionary perspective. Huge congratulations to the entire team on this monumental launch. The release of Atlas sets an entirely new benchmark for spatial intelligence and world models. Thank you both so much for sharing your insights today!