Catching Heat Technical Demo

Catching Heat 0.0.20 Final Technical Demo released!

Disclaimer: These are alpha-quality builds. Our AI artist is still in learning.  Do not expect visual quality.

This is the final demo version of the game. We have confidence now we can move forward with the technology we have developed for almost a year. Please give it a spin and let us know what you like, what you dislike and how we can improve. If you can, consider supporting the game as we will need additional support as production ramps up. We have a wish list of things that will make development a lot faster.

Check out the release notes and screen shots for 0.0.20 and take note: First public drop of the next version 0.1.0 will be October 30th with monthly releases after that. Patreons get access to weekly builds and even customized builds (at some tiers).

The Journey so far…

Catching Heat 0.0.1 was released May 15, 2020, but the technology has been in development since December of last year.

“Gen” (Generated image) from Catching Heat 0.0.1

Earlier in the year, we engaged in much research and experimentation with different networks and learned a lot about what works, but mostly about what doesn’t work. It was very exciting to wake up every morning and see what the AI models created over night. After months and months of training, we found some networks that produced far better results than our earlier experiments.

sample from failed network training experiment

The network that we ended up using is called Pix2PixHD. It’s able to produce HD images but was lacking in the ability to control the attributes of the output images. Our early versions had no control over hair color, clothing or anything else but the pose. We had to modify the pix2pixHD network with this capability and we call this pix2pixHD_ACM which is now open source. It wasn’t until June, after months of training and retraining this new network that we reasonable results from it.

New pix2pixHD_ACM generated image from 0.0.7

There were still some issue with the new network, and it took several architecture changes to get memory usage down. Our 4th generation model was taking over 9GB of GPU memory to generate the images. We also wanted to change the architecture in a way that we could train different parts of the network across different machines. Thus generation 5 of the network was born. This architecture enabled us to train the now production model up to decent quality in 3-4 weeks and on only 20k images that we created from scratch for the game. The last release using the 4th generation model, which trained for nearly 3 months straight, was 0.0.15, one of our best-ever builds.

The girls from build 0.0.15

A new dataset and a new model

For a patreon-exclusive series of builds, we created a dataset of 20,000 images using Wildeerstudio’s “Lara” model. We modified the clothing, hair and poses and rendered countless images. It started out as an experiment in “how good can quality get if we train on a single character” but the results quickly became more and more impressive:

Soon after, we discovered a data augmentation technique to force the network to generalize better. Data augmentation basically ensures that the network never sees the exact same image in the same way. This is often done by resizing and cropping and we used this heavily over the advancing weeks to improve the model. We often have the model generate a video translation to see how well it is generalizing and able to adapt to poses the network has never seen (first video is 0.0.15 gen4 model, second is new gen5 model):

Fast forward an additional month and the results were looking really good:

The results were so good compared to previous attempts it was an easy decision to start using this model to generate game art moving forward starting in build 0.0.19.

The addition of voice

Build 0.0.19 also included the first AI voice powered by the wonderful and famous talents of MidnightDatura. After a week of training, the little voice critter we call “DaturaNet” was progressing well. As of 0.0.20, it still isn’t up to the quality we expect, but she’ll get there. It took us humans several years to get speaking down and DaturaNet has only been training for a few weeks!

We want to feature full voice acting for the game. Euryale is spoken for. Who would you like us to grow AI voice critters after next?

Where do we go from here?

We have outlined about 46 scenes to complete Catching Heat. We will start adding new scenes weekly for patreon supporters and monthly for the public. Depending on funding, we will add additional voices and AI generated music to the game and also invest more in our training dataset which has grown over 30k images.

It’s been an incredible journey so far developing this game and I (deepglugs) hope that you enjoy one of the first AI generated adult game ever. In the very least I hope you enjoy being shocked at the weird stuff AI creates. This game should be one of the most unique experiences you’ve had playing a game so far.

Notify of
Inline Feedbacks
View all comments