Okay, so yesterday I was messing around with some face generation stuff, specifically looking at how to create images that resemble… well, a certain style. I stumbled upon some research mentioning a name, “jaime lowe,” which apparently is linked to some specific aesthetic qualities in generated faces. Thought, “Hey, why not give it a shot and see what happens?”
First thing I did was dive into the research papers I could find that mentioned the name. Most were about GANs (Generative Adversarial Networks) and how they could be tweaked to produce specific types of images. I wasn’t trying to build a GAN from scratch – ain’t nobody got time for that! – so I looked for pre-trained models I could play with.
I found a couple of StyleGAN2 implementations on GitHub. StyleGAN2 is a pretty popular architecture for generating realistic-looking faces. Grabbed one that seemed reasonably well-maintained and had good documentation. It was a Pytorch implementation, which is what I’m most comfortable with. Then cloned the repo and started setting up my environment.
Got my conda environment sorted with all the necessary dependencies – Pytorch, CUDA drivers (important for GPU acceleration!), some image processing libraries like PIL, and a few others. This step ALWAYS takes longer than I expect. There’s ALWAYS some weird dependency conflict or version mismatch. Eventually, got everything working. Did a quick test run to make sure the basic face generation was working, and it was – glorious, slightly creepy, randomly generated faces.
Now, the tricky part. I didn’t have a specific dataset labeled “jaime lowe faces,” obviously. What I DID was look at images associated with the name online. Gathered a bunch of these images manually, focusing on those that seemed to capture the specific look I was going for. It wasn’t about replicating any single person, but capturing a certain style.
Then came the data preparation. This was a pain. All the images were different sizes and aspect ratios. I wrote a quick script using PIL to resize and crop them all to a consistent size (512×512 pixels in this case, which the StyleGAN2 model expected). Also normalized the pixel values to be between -1 and 1, which is a common practice in GAN training.
Next, I needed to “fine-tune” the pre-trained StyleGAN2 model with my “jaime lowe” inspired dataset. The repo I was using had a decent fine-tuning script, so I adapted it to my specific dataset and training setup. I didn’t have a massive dataset, so I kept the training relatively short to avoid overfitting – didn’t want the model to memorize the training images! I monitored the training loss to make sure it was decreasing, which meant the model was actually learning something.
After a few hours of training (thank you, GPU!), I had a fine-tuned model. Time to generate some faces! I used the model’s inference script to create a batch of images. The results were… interesting. Some were pretty close to what I was aiming for, while others were just weird and distorted. It definitely picked up on some of the key features from my dataset.
Finally, I did some cherry-picking. Generated a bunch of images and selected the ones that looked the best. Used some basic image editing tools (Photoshop, GIMP, whatever) to clean them up a bit – removing any obvious artifacts or distortions. It’s a bit of an art, not a science, this part.
So yeah, that was my little experiment. It wasn’t perfect, but it was fun to see how I could influence the output of a GAN by fine-tuning it with a targeted dataset. And now I have a folder full of oddly specific, AI-generated faces.
