Raymarching Cloudscapes

https://www.youtube.com/watch?v=WHOPX68eTzk&t=40s

Terrain shown in the video is from https://github.com/TokisanGames/Terrain3D

This repository contains the results of my exploration and experimentation with creating realistic cloudscapes using raymarching techniques. Raymarching is a rendering technique to create 3D computer graphics. It's especially useful for rendering volumetric and procedural scenes. While the primary focus is on volumetric clouds, I've also incorporated other effects for a more complete skybox look.

Project Overview

• Objective: Learn and experiment with raymarching to create dynamic and visually appealing cloud formations.
• Note: The project started in raw glsl but eventually it was moved to Godot in order to facilitate testing. The focus was not to exploit Godot's capabilities and some more engine niceties could have probably been used. Keep in mind that this is not meant for use and there are many things that could be improved, there are a few that I am already aware of but that was the main goal, to be aware of certain things that I was not at the time of starting.
• Next Steps: Although this repository serves as a learning experiment, in the future I may create a separate, more specialized repository for a fully developed clouds and weather system, also including a physically based atmosphere. It will not necessarily be in Godot and it will probably be in a custom Vulkan based render.

Some Optimizations and Techniques Used

1. Triangle mesh in Front of the Camera: The raymarching is done in a triangle mesh in front of the camera with a pixel shader. I utilize the z-buffer to render on top of the scene or skip rendering clouds altogether to eliminate unnecessary calculations.

2. Stepping Techniques Tested:

   • Dividing the step size based on distance and view angle: Fewer steps are needed when looking straight up to achieve good looking results due to how we represent the atmosphere and the position.
   • Marching at increasingly larger steps with various variation functions (linear and non-linear) up to a maximum distance.
   • Marching at larger steps initially and refining with smaller steps once a volume is found and going back to the other method once we failed in detecting a volume a certain amount of times (inspired by the Horizon Zero Dawn technique).

3. Noise Construction:

   • A 2D map defines cloud surface existence. This allows artist authored scenaries or some simple generated noise texture. Otherwise some compute weather system could be set in place for more dynamic cloud changes over time.
   • Smoothmin combined with three layers of pregenerated noise textures (Perlin and Worley).
   • Detail noise as erosion from bordera, selectively applied based on transmittance thresholds or in the last stepping method described when looking for volumes. This can be skipped based based on a threshold when transsmittance is slow to save us some time.

4. Flat Distant Clouds: To simulate distant clouds, a 3D texture is mapped onto the desired sphere radius and a Y scroll is used to simulate changes over time (a 2D texture could also be used).

5. Dithering and Anti-Aliasing: Dithering with a scrolling blue noise texture or TAA are used to reduce the number of steps needed to maintain quality results.

6. Ray Sphere Intersection: In order to render the clouds correctly at each point the intersection with two different sphere had to be calculated (inner sphere and outer sphere), and used starting point and end point of the raymarch in order to skip stepping over empty space. Some cases were no intersections are found also had to be taken into account.

7. Specific to Godot and this project: I have used a custom helper addon of my creation Noise Texture Composer, in order to combine noise textures into a single one to reduce the amount of texture samples needed in the ray marching loop.

There are probably more things that I have forgotten to mention.

Results

• On a 3440x1440 monitor with an RADEON RX 7700 XT graphics card, 2.8ms render time could be achieved with relative ease with various different uniform setting, although it generally had to be at the cost of rendering distance or looks. Due to the project being for experimentation purposes there was no goal to achieve better performance for a specific use case. Some dialing would need to be done to satiay specific project needs.

Sources and Inspirations

• Sebastian Lague's Coding Adventure: Clouds
• Simondev's How Big Budget AAA Games Render Clouds
• Cloudscapes in Horizon Zero Dawn
• Shadertoy community
• Principles of physically based rendering

Feel free to explore the code and continue experimenting.