OpenGeofiction

For people who want to make unique pictures of landscape. (Note: Not advertisement)

Posted by theruler on 23 December 2019 in English (English)

There are probably some people who have heard of GauGAN, a program developed by NVIDIA.

It is an AI-powered engine which makes landscape, of your choice, based on millions of photos of geography.

It works like paint, you paint some things on it, click on a button, and boom it becomes a landscape.

It is an online program.

Right now, GauGAN is open for demo.

This means that y'all have access to making unique landscape with tools such as clouds, mountains, hills, water and rivers, to perfectly match your mapping.

Note that GauGAN is AI, so it may be confused at some things you draw, for example if you detail it too much.

Note that this is not an advertisement, but rather a reminder for people wanting to create unique landscape eventually for a wiki article.

GauGAN Demo

Demonstration

I actually made this landscape. This is a demo for some local hills and mountains.

Comment from Arlo on 24 December 2019 at 03:52

Wow! Thanks for sharing! This is a great resource.

Hide this comment

Comment from SwissCrusader on 26 December 2019 at 00:50

Wow, that's pretty good for a demo software. Looks promising.

Hide this comment

Comment from niski on 26 December 2019 at 10:35

using this we can make OGF Street View xD

Hide this comment

Comment from MrManny on 6 January 2020 at 10:42

I've heard of it and I think it's super great. But I will hijack this real quick and openly wonder if we can find (or train) a GAN for satellite images. I'll have a look into that and check, if someone has already done something like it. Unfortunately, my machine learning skills are not necessarily good enough to do something like that from scratch :-(

Hide this comment

Comment from niski on 6 January 2020 at 11:31

As i know, there's not a public GAN which could do that. I might try learning how to code a GAN, if we get something, we could add a new layer to the map!

Hide this comment

Comment from niski on 8 May 2020 at 21:27

old topic, but i recently built my own GAN based on the same resources NVIDIA GauGAN uses. it is trained on satellite imagery, though it doesn't have enough data and training. once i am satisfied enough with the results i may post them in a separate diary entry

Hide this comment

Comment from MrManny on 9 May 2020 at 10:38

@niski: that's exciting news! keep us posted! :) and let me know if I can be of help. My experience with GANs is very limited (practically zero), but other than that I do believe I am a passable software engineer.

I think one of the bigger challenges is the sparsity of regions with a lot of detail on OGF. OSM and the likes tend to have very detailed maps (e.g. buildings, individual trees, etc.) due to manpower and pre-existing data sources. But maybe a GAN-generated satellite image layer would motivate people to also map exactly that.

Hide this comment

Comment from niski on 9 May 2020 at 10:43

as of now, it can generate fairly mediocore stuff from sketches, maybe i'll yet add some data to train on and train it even more.

important thing to note is that unlike GauGAN, the more detail the better the result. the biggest con right now is the fact the GAN still has troubles with generating good-looking residental area. Due to mask limitations i had to merge several things such as city area and residental area, though for smaller towns you can use regular buildings and it does the job better.

Hide this comment

Comment from niski on 9 May 2020 at 10:49

@MrManny referring to your comment

  1. help is apperciated, especially adding more data (i will make some sort of a tutorial how to do it properly)

  2. lack of detail will probably result in more wonky images, i tried comparing the images with less and more detail and the ones with more detail simply had less noticeable bugs/inaccuracies, making the image nicer. i will likely test the GAN on several OGF images and see the diffrent results

Hide this comment

Comment from niski on 12 May 2020 at 07:16

the AI has been tested and is trained enough to make some basic satellite images based on OGF map (and other drawins, lol i literally made a forest shaped as some random character with lakes inside and it still looks fine). not sure how i can release it publicly yet, as i had to host it on google colaboratory and you can't seem to remotely access it from ex. another website

but if you want you can message me with a location and i can post the result (note: not all of it will be taken properly as it has to be a zoom at least 500m, and the AI only makes 512x512 images anyway)

Hide this comment

Leave a comment

Parsed with Markdown

  • Headings

    # Heading
    ## Subheading

  • Unordered list

    * First item
    * Second item

  • Ordered list

    1. First item
    2. Second item

  • Link

    [Text](URL)
  • Image

    ![Alt text](URL)

Login to leave a comment