I'm uploading an alternate model that feels more accurate, in my opinion. I've altered my working process a little bit in the last few days. First of all, I don't use the "Use PNG alpha channel as loss weight" anymore, as I felt my embeddings were getting overtrained almost from the start. I have a more active approach now too: I begin with a flat learning rate of around 0.004 and patiently wait until I get some decent pic resembling the subject, which usually doesn't take more than 400 steps anyway. Then I interrupt the process, test the result and reduce the learning rate to maybe 0.001. Run the process for a few more hundred steps and usually I already get a decent TI at least. I guess I'll keep improving the process as I run more tests, but for the time being I'm happy with these results. Hope you like it!