I was looking for algorithms that quickly compute quite nice realistic, or not realistic, animated landscapes, in video format ("animated" = the camera moves, or the water is animated, etc.) But all I found were: deep learning GANs, which are difficult to train, or even difficult to set or transfer. ("hard" and time-consuming too)

The animated landscapes rendered can be like children's cartoons: I just want to quickly generate landscapes, which have to be quite beautiful, that's all.

What research papers, documentation, tutorials, tools, APIs, libraries, Github/Gitlab repositories should I read? MIT license or equivalent please.

Thank you!

  • $\begingroup$ Beautiful being subjective, I will focus on the objective part of generating any plausible landscapes. I suggest looking into procedural terrain generation. I am referring mainly to using some noise (value/perlin/worley) and then layering it at different scales (fractal brownian motion). You can find many examples on shadertoy. $\endgroup$
    – lightxbulb
    Commented May 9, 2020 at 5:10
  • $\begingroup$ Thank you I will look on Github and HackerNews for papers/tutorials/sources examples. Do you think I can generate terrains like these ones (less or not realistic of course, with more abstract or gaming textures than these realistic ones): youtu.be/ceGLEhahLKQ?t=1 ? $\endgroup$ Commented May 9, 2020 at 12:27
  • $\begingroup$ Generating terrains as in the video you linked may be feasible, but it's certainly not easy. Here's a representative example of something considered quite good: shadertoy.com/view/4ttSWf $\endgroup$
    – lightxbulb
    Commented May 9, 2020 at 13:14
  • $\begingroup$ OK thank you! Well, I have developed a procedural terrain simulator. When I launch it, I end with a 3D terrain with water, a "plausible terrain". But it's not beautiful: there is no snow, no grass, no trees, no flowers, no... nothing, just the terrain and the water. Di you think it would be feasible to use a deep-learning model "Picture2Picture" in order to add these textures, these things to my 3D virgin terrain image? Are "Picture2Picture" networks easier and faster to train than GAN (less than 1 hour)? Do you know if there is a pre-trained Pic2Pic model that could help me with my 3Dterrain? $\endgroup$ Commented May 9, 2020 at 13:33
  • $\begingroup$ I have no idea on how to use ML for this task. As for formulating this as an answer - I believe others may have more valuable input on this. $\endgroup$
    – lightxbulb
    Commented May 9, 2020 at 14:59


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.