Disney's AI Learns To Render Clouds | Two Minute Papers #204

TL;DR

Disney’s hybrid cloud-rendering method uses a neural network to predict in-scattered radiance while traditional algorithms handle other parts of the renderer. Trained on 75 procedurally generated and artist-drawn clouds, it reduces rendering from as much as 30 hours to seconds or minutes. The technique also enables interactive parameter editing and temporally stable, flicker-free animation, so read on to see how it improves both speed and creative control.

Transcript

Dear Fellow Scholars, this is Two Minute Papers with Károly Zsolnai-Fehér. This is a fully in-house Disney paper on how to teach a neural network to capture the appearance of clouds. This topic is one of my absolute favorites because it is in the intersection of the two topics I love most - computer graphics and machine learning. Hell yeah! General... Read More

Key Insights

• ⛈️ Rendering realistic clouds involves volumetric path tracing, which requires simulating millions of light paths with scattering events.
• 😶‍🌫️ The traditional rendering approach can take up to 30 hours to render an image of bright clouds.
• ⌛ The proposed neural network approach significantly reduces rendering times to seconds or minutes by learning in-scattered radiance.
• 😶‍🌫️ The neural network is trained on a dataset of 75 clouds to capture a wide variety of cases.
• 👻 The technique allows for interactive editing of scattering parameters without the need for lengthy trial and error phases.
• ❓ The rendered images using deep scattering are indistinguishable from reality.
• 🥶 The technique is temporally stable, ensuring flicker-free animation rendering.

Questions & Answers

Q: How does Disney’s AI render clouds faster?

The neural network learns to predict in-scattered radiance, so the renderer does not have to calculate that part itself. Other parts still use traditional algorithms, making the method a hybrid approach that renders images in seconds to minutes instead of many hours.

Q: Why are realistic clouds difficult to render?

Realistic cloud rendering requires volumetric path tracing, where rays can penetrate surfaces and undergo many scattering events. Simulating millions of light paths with potentially hundreds of scattering events is computationally demanding.

Q: How long can traditional cloud rendering take?

A bright-cloud image can take up to 30 hours to render with the traditional approach described. Increasing the number of simulated scattering events improves realism but also increases rendering time.

Q: How was Disney’s cloud-rendering neural network trained?

The training dataset contains 75 different clouds. Some were procedurally generated by a computer, while others were drawn by artists to expose the learning algorithm to a wide variety of cases.

Q: What is in-scattered radiance in Disney’s rendering approach?

In-scattered radiance is the concept that the neural network learns to predict rapidly. By predicting it, the system avoids directly computing that portion of the rendering process.

Q: Can artists edit cloud scattering parameters interactively?

Yes, the scattering parameters can be edited interactively without waiting hours to evaluate each new setting. This makes the normally lengthy trial-and-error phase almost instantaneous.

Q: Does the deep scattering technique support different scattering models?

Yes, the technique supports a variety of different scattering models. This allows it to handle different ways of capturing the scattering behavior that shapes a cloud’s appearance.

Q: Can Disney’s deep scattering method produce stable cloud animations?

Yes, the technique is temporally stable, allowing animations to render without flicker. Its results also appear mostly indistinguishable from the ground-truth renderings shown in the comparison.

Summary & Key Takeaways

• Rendering clouds using light simulation programs requires volumetric path tracing, which involves simulating rays of light that penetrate surfaces and undergo scattering events.

• This computationally demanding process can take up to 30 hours to render an image of bright clouds.

• The paper proposes a hybrid approach where a neural network learns in-scattered radiance, significantly reducing rendering times to a matter of seconds.