When we say Atlas has pixel-perfect camera control, we mean it. Atlas precisely follows input camera parameters, including non-planar projections such as the Brown-Conrady distortion model and the Kannala-Brandt fish-eye model. Panini Bhamidipati invented a novel method of camera conditioning and it works beautifully. 🧵 [1/N]

Pinhole cameras are a given -- Atlas can perform the iconic dolly zoom effortlessly.

Non-planar projects are also no problem. Here's a fish-eye projection with the Kannala-Brandt model.

And pincushion, barrel, and tangential distortion under the Brown-Conrady model.

Read more about Atlas: Request early access:

Thank you! It was a very fun problem to work on with this incredible team.

@ychngji6 @BhamidipatiPan1 Interesting! So can yall reconstruct a scene given a sparse set of say 8mm fish eye images taken on a full frame sensor?

@ychngji6 @BhamidipatiPan1 8mm might stretch the current model a bit, but in principle yes.

@ychngji6 @BhamidipatiPan1 Very cool

following the camera parameters is the easy half, the harder one is whether the reconstructed geometry stays put once you move past a clean single-subject shot. cluttered scenes with occlusion and reflective surfaces are where world models usually drift, curious if that holds at this precision

@BhamidipatiPan1 Does this qualify?

@BhamidipatiPan1 Good stress test, plenty of clutter and occlusion in there. Curious if it still holds on the reflective tool surfaces, that's usually where the drift shows up first, not the diffuse stuff.

@BhamidipatiPan1 Super nice job! Thanks for sharing! Have you consider adding more characteristics like bokeh / aperture and shutter speed (motion blur), which are also very physical and optical.

@BhamidipatiPan1 Not yet, but good idea!

@BhamidipatiPan1 Very cool. On uncalibrated input, ballpark how tight are the recovered intrinsics and distortion vs a ChArUco calibration? Sub-pixel reprojection, or more like a few px?

@BhamidipatiPan1 I trust camera control after I check the generated sequence, not the still; continuity has to survive the path users actually hit.

@BhamidipatiPan1 Atlas pulling off pixel‑perfect camera control is wild engineering. When you can model distortion this precisely, you’re not just simulating reality, you’re bending it. Thoughts?

@BhamidipatiPan1 Good stress test, plenty of clutter and occlusion in there. Curious if it still holds on the reflective tool surfaces, that's usually where the drift shows up first, not the diffuse stuff.

@BhamidipatiPan1 What units are used in Atlas? One thing I hate in blender is figuring out how large to make something compared to my scene

@BhamidipatiPan1 Whatever units you want. It's grounded to your inputs

@BhamidipatiPan1 I am loving all the demo twitter videos - anyway I could get access?

@BhamidipatiPan1 Ray embeddings?

@BhamidipatiPan1 Most camera-conditioned models fake it with a pinhole assumption, so bothering to nail actual fisheye distortion is a genuine technical achievement.

@BhamidipatiPan1 Camera control is what makes it usable for real shots. Where does it still drift, long moves or fast ones?

@BhamidipatiPan1 👍👍