At the encouragement of @PaulBellow I decided to write a dev log series on the AI Game Studio and the game I am currently developing (2.5D interplanetary cyberpunk action-adventure). I am building it with Codex so it’s highly relevant here. The point of this dev log is two-fold:
Share my learnings so you can apply them for your own fun and profit
Get feedback from the community so I can improve, learn, and have fun doing it!
Before I get into the details I’d like to point out that the v0.2.1 of the AI Game Studio allows you to create projects consisting of sprites/animations and sounds. The sprites are automatically saved in both .png and .aseprite format so you can load them natively in Aseprite for editing.
The way I am building my game AFTER LIGHT is I am using the game studio to create sprite and sound assets (both soundtracks and sfx) and then drive the overall game creation via Codex. Here is the high level flow:
It’s a 2.5D game where the levels/world is built in 3D while characters/sprites are pure 2D pixelart created with my game studio
For 3D models in Blender I try to keep them as simple low-poly shapes and use Astra for modelling
I place heavy emphasis on pixelart textures, volumetric lighting, depth of field, and blur, rather than complex 3D models
I use imagegen skill for generating the textures
I tend to switch between Astra and Sol models, depending on the task
Now some learnings when developing the AI Game Studio, which I am also building with Codex:
Initially I tried creating spritesheets with pure image generation - both single-shot (generating a PNG with rows and columns) and conditional generation - using the previous image to generate the next one; both approaches resulted in heavy misalignment and quirky movements. I am an amateur artist and animator (both analogue and digital) and I have a high bar for animations.
What worked for me instead is using GPT Image models for creating a reference pixelart character, and then using a video gen model (like MiniMax H3 Max) for creating a short action sequence (3-5 seconds). I then use ffmpeg to extract and resize the frames (e.g. in 128x128px); I then perform simple de-selection of frames to cut it down to 12-24 frames per second. All of this is done very easily in my game studio.
Even though GPT-Image-2.5 has amazing support for transparency, I still have to apply chroma keying techniques due to video generation.
I often need to load the sprites in Aseprite for some touch-ups and inconsistencies. If you look at the gameplay gif, you will notice “John” wearing an ID badge - sometimes this badge disappears and I have to add it manually.
For soundtracks and sfx creation I use ElevenLabs Sound Effects API - their free tier gives you 10k credits per month which is plenty for this type of work (you will need at least a starter pack if you want to go commercial).
All the other models (GPT Image and MiniMax H3) I use via OpenRouter.
So I pushed a new release and in this one I did a couple of changes that I feel are just a big quality of life improvement when generating characters and animations.
First off, when generating characters it’s great to have multiple aligned views of the same character. In my case I generate side, front, and back. This gives me consistent reference images to condition on, and create sequences from various angles. In the past I always had issues with “baseline alignment”, but with GPT Image 2.5, this seems to be largely solved. Note that this is not mandatory, and i was able to generate front and back animations based on the side profile only, but then you are leaving things up to change, and there are always some issues. For example, the ID badge disappears, or the colour of the hair changes slightly. So having multiple frames of reference just gives you that consistent generation across different angles.
The second quality of life improvement is conditioning on start/end frames when generating animations. For example, if you want a perfect looping idle sequence, you want your first and last frame to converge. Similarly if your character is performing a sequence where she turns north-bound, then returns back to side-view, you want that consistency on the last frame.
This is a small update, no new functionality, but more about the “packaging”.
Firstly - and this is such a meta thing - I used the AI game studio to build the “mascot” for the AI game studio. Introducing…Wombo! Keeping in line with the Australian animals, he is a personification of an Australian wombat.
I plan to feature Wombo in little “waiting” and “loading” sequences in the game studio.
Also - I introduced some GIFs showing examples of usage. This is one thing I personally appreciate when visiting peoples’ GitHub repositories - seeing images or GIFs of the product in action.
So I released a playable demo for AFTERLIGHT. To complete the demo you have to demonstrate some basic Linux admin skills (hint: think file ownership and read permissions). It’s nice to see all those assets and work in Codex come together in a single tiny playable package.
One thing to note is that sprite/animation generation is not perfect. Even though I apply chroma keying, some green is not fully filtered out as you can see in this screenshot. Ultimately, to make this fully polished you have to load the sprite in Aseprite and do some touch ups. The result is only as good as the quantity of love you put in.
Added some more useful features. First off, if you are like me, you tend to sketch things out with a pen on whatever piece of paper you have close by. Sometimes when I am on the metro I might get really inspired by someone’s look and sketch out a character real quick. Or you might just have some reference image that you find very inspirational. So now you can upload that image and combine it with a prompt and get the character you want.
Another useful feature is to test your sprite and animation against different palettes. For example, you may want to see how your sprite walks in a glowing, light, or dark backgrounds. It’s also a great tool for seeing if there are any colour discrepancies due to chroma key removal. Then you can do some final touch-ups in Aseprite.
I also made some UI consistency improvements. As good as Astra is, it doesn’t care about enforcing consistency among UI placements, naming conventions etc. For example, I want a particular order of buttons to be always the same no matter what section of Wombo I am working in; likewise I want the same naming conventions to apply everywhere. So you have to be explicit with this and Astra will do you proud.
Some extra features to bring more professionalism (and usefulness) to Wombo.
I created a prompt library, or more specifically, a style library, for character creation. You can now choose between 8-bit pixelart, “cinematic” pixelart, 2D cel style, or 3D “storybook” style. Or you can just go with your own custom style. I also fixed a bug - when generating animations, it biased towards pixelart style, but now the character style is consistently propagated right through.
There are also some small fixes on UI and copy. When you are creating applications with AI, you have to be extra careful how things are named, and the UI behaviours. Does hitting ESC or ENTER do what is intended?
