Testing Faceswaps with ChatGPT...

 

I'm back already. I was testing around with the swapping capabilities of ChatGPT, and I figured they turned out well enough to publish as official works-especially since I only had one fake for the last post. Speaking of which, I mentioned that I was using ChatGPT for the headshot of the last post, the reason is because I have canceled my Gemini subscription for now. It was so buggy and inconsistent that I wasn't able to generate anything. It would either say it "couldn't depict certain public figures" or "I am unable to fulfill the request at this time." In regards to the first one, it was triggering that with people that didn't even exist or even animals. Needless to say, I had enough of its bullshit, so I decided to go back to GPT to see how far its image generation has come, and use the Gemini that is integrated into Photoshop proper if I need to. It can be hit or miss with GPT, but when its a hit, its a hit. 

So my test subjects of choice for this set are my beloved Willa Holland (I had to, sorry), Chloe Grace Moretz, Emma Myers, and Freya Allen. It terms of a hit, I think Willa, Emma, and Freya are fantastic. They have a high quality detail and realism to them that rivals what Gemini is capable of. Meanwhile, I have to admit that Chloe is a bit of a miss-as it doesn't contain the same qualities and looks more like the typical "AI slop" that circulates the internet like wildfire (though with much more effort put into it). Still, I figured some of you will be satisfied with it enough. 

I think this method of swapping has its uses. In particular, it can help in removing the hair from the original subject to account for the hairstyle of the new subject (Willa and Emma are the clear examples of this). You would think that would be easy enough to do in Photoshop, but not in my experience. Also, it will generally try to match the skin tone and lighting on its own-giving me much less work to do overall. The disadvantages this method has is always in overcoming some hurdles. For obvious reasons, I cannot work with the full image in the swap. I have to crop down the image enough to where it will work with it. Because of this, the generated image does not align properly with the original when you bring the two together in Photoshop. Thankfully, the process of aligning things properly isn't too difficult. You need to remove the background, align the head properly, bring some of the background back to cover parts of the original subject, make any edits necessary to match skin tones, and use the remove tool wherever its needed around the hair, head, neck, and background. Once all that is done, it is smooth sailing from there.

I will testing with ChatGPT further, but I feel that I will return to Gemini at some point in the future. For now, I have a few ideas for my next works: a Resident Evil 2 fake, and a couple Dragon Age Origins fakes. I already have the RE2 fake started, and I have the materials ready for the Dragon Age fakes. I think GPT will help tremendously in bringing those fakes to life, but until then, enjoy these fakes, and stay safe out there.