Stable diffusion gbafegirl gbafeboy on google colab

Stable diffusion gbafegirl gbafeboy on google colab

I created a LoRA that outputs portraits like GBAFE

I ported google colab which is available for free.
However, webui is a pain to deal with LoRA, so it is run from the command line (google colab ui).
There is no tag completion, etc., but please be patient with that.

Must have a google account.

Access the following google colab

Press the “Setup” play button to set up the VM.
It will take a few minutes to set up several GB of data.

Press the “DrawTest” play button.
It will take 1-2 minutes to complete the generation process.

Each time you press the Play button in DrawTest, 10 images and a zip file containing them are created in /content/tmp.
The list of files can be seen in the left panel.
Each time an image is created, all of them are deleted, so it is a good idea to download them when they are generated.

The free version of google colab only allows you to use it for 2-3 hours a day.
If you run out of time, use another google acount or charge 10$ or try the next day.

If you want to generate BOY’s, change network_weights.
network_weights = “/content/dreambooth/output/gbafegirl.safetensors”
network_weights = “/content/dreambooth/output/gbafeboy.safetensors”

The default prompts are as follows
prompt = “green background, 1girl, open eyes, blue hair, long hair”

Try different things.
prompt = “green background, 1girl, open eyes, red hair, short hair”
prompt = “green background, 1girl, open eyes, open mouth, green hair, ponytail”
prompt = “green background, 1girl, open eyes, pink hair, twintails”

If you create a nice picture, please share it with us.
If you process it into Portrait format, it would be very nice if you could post it in the repository!


Cool novelty I guess.
Not creating actual pixelart so the usefulness of such generator is limited. Might be good at providing some inspiration for design, though.

1 Like

Been doing prompt for a couple hours and it’s cool, but it’s not exactly there quiet yet. Good for concepts, but sadly falls short of creating insertable portraits.

These are the 3 best ones I was able to get. A very good start.


That second one looks suspiciously familiar. Hmmmm.

I’ve also been playing around with these for a bit, I’ll probably toss a few up when I’ve time to sort through them. I might be trying to resprite a few whenever I’ve more time for fun.

But seriously, they were all prompted to be Eirika. Just got some lucky rolls.
Wanted to make a 96x80 but…
This gremlin popped up.


Well a few of my favourite ones so far. I’ll probably try to do something with them if I can unlazy myself.
The lower one in particular for 6, I rather like how it kinda just ‘looks’ like a portrait without actually being one, haha.

1 Like

Try cropping them to the aspect ratio of 5.1x4.2, downscale it to 96x80, quantize/reduce color depth to 16, and drawing over it then making any necessary fixes.

1 Like

That is something I am not capable of doing, simply because I wouldn’t know where to start

some are weird, some are very eliwood, but uh

is this a watermark/what is this trained on


A few more

These are actually really fun, thanks for the tool!

PaintDotNet is the best application for spriting in my experience. You can grab a sprite sheet from spriter’s resource for palettes, just take it slow, use the pencil and start with an outline, and do your shading from the darkest shade up.

Proof of concept for what I am suggesting.

It’s not a perfect recreation of the art, downscaling muddles the shading but I did my best to improvise.

Needs a little more polish but it’s late and I’m tired. lol


honestly looks like a combination of gacha game art and vanilla gba portraits which has me lolling

…this is crazy.
this thing will probably make splicing outdated by the time it gets improved, AI are all over the place lately and they’ll progress even further as time passes.

anyone using Photoshop or other professional graphic softwares can just generate an image, make fixes where needed and voilà, you pretty much get a free portrait with low effort.
i’ll have to look into it too, eventually.

1 Like

This letter is not in the data trained by LoRA, so I assume it is something in the Anything3.1 model.
What prompt did you use to output that image?

When I mistakenly forgot to put a green background, an interesting picture was output.
She looks like the female boss of the enemy.
It still needs some cleaning, but I got it to work for now.

1 Like

Pretty sure it was “1man”, “black hair”, “short hair” with feboy

Followed your steps:



I didn’t requantize and I didn’t spend much time on it, but I’m 100% certain this is doable. Thanks for sharing your method!

Not a problem, as long as it inserts and plays smoothly that’s what matters. Plus depending on your emulator you can set it to have that low res blur like a real Gameboy Advance and that makes a lot of chop unnoticeable.


After a little bit of work, I think I got one insertable. No actual frames or anything 'cause this is mostly just proof of concept in the insertable format.
Full Portrait Trial 1
A couple of things working on this one taught me:

  1. Not all generated images make good portraits. Plenty were either zoomed too far out so that lower pixel counts completely ruined them, or so far in that parts of the subject extended past the edge even in the original image (usually the shoulder and hair/back of the head). This could be mitigated, but not entirely removed, by adjusting scale and image size.
  2. The colors most of these images use are too colorless. I turned the saturation on this one, and on plenty preliminary attempts, to 120 and it works much better.
  3. These images are usually looking directly at the camera, and a number of them, including this one, are facing the camera as well. Haven’t found a way to mitigate, may just have to manually edit. (just a thought I had 30 seconds afterwards: changing the prompt to have some variant of “facing sideways” might work)