<p>Version 2 of this textual inversion renders Debbie Harry of Blondie pretty well. It is trained on photos of her from the 1977 - 1982 period. This model may not be used commercially. If you download the training data file, it has textual inversion files for every 10 training steps. I found that 3500 training steps was best (so it is the default), but often the 4500 training step model does better. </p>

Debbie Harry from Blondie

<p>Debbie Harry has one of the world's most uniquely symmetrical and well proportioned faces, which means that even small errors are really noticeable. As it happens, even a difference in a couple of hundred training iterations makes a big difference in how the images come out. I've packed all of the files, including the training checkpoints and different inversions with different numbers of training steps into a zip file that is attached as the "Training Data File". The main model used 3500 steps, but in many cases other numbers of steps give better results. The "training data" includes every number of training iterations, in steps of 10, from 10 to 4500. A few of the sample images show the same prompt as rendered on a textual inversion with a variety of iterations. If you install more than one of those other inversions, the tag would be "debbie_harry-####" where you replace #### with the number of inversions (so the 4500 inversion one would be debbie_harry-4500) I learned a lot doing this training, so hopefully my future efforts will be improved as well.</p>

Files with SHA256: a4e088bb802c91bbaa37e64ff96929f01220d10c8e4ccd529395d4c0a99256d4