<p>Work in progress. i was really dissatisfied with the current models for pirates.  The inferences always seemed to include a comic skull and crossbones painted on a hat.  And the tricorn hats always looked wrong. So I am attempting to make a more aesthetically rewarding models for pirates.  I posting four versions along with their datasets.  I hope someone can give me some guidance on how to better elaborate details like the shape of a cutlass, the shape of a tricorn or bicorn hat. </p><h2 id="heading-11">Version 1  </h2><p>First out the of the gate.  Came out a little too Johnny Deppy, if you know what I mean.  Still tends to consistently deliver credible looking pirate captains, mates and women pirates. Made a mistake and baked this with a dimension of 9 and an alpha of 1.  But I like the results. I'm going to experiment with using lower dimensions and alphas in the future. Works with the trigger keyword p1r4t3s, but you can omit it for different effects.  The other keywords are "captain", "mate", and "female". Don't use the word "pirate" in your prompt or you will get a bunch of dorky cheesy pirate costume dreck that is in the base models.</p><h2 id="heading-12">Version 2 </h2><p>Jacked up the network alpha and dimensions to 128, but didn't make many changes to the captions or weighting in the data set. Same trigger and keywords as the first version. Aslo changed clip skip to 2, just to see what happens. </p><h2 id="heading-13">Version 3</h2><p>Lowered the network dimension and alpha to 16.  Changed clip skip to 2. I did a lot of reweighting of the data set so that Johnny Depp was weighted a lot less and female images were weighted more, especially the character of Anne Bonny from Black Sails. This improved the originality of the images produced.</p><h2 id="heading-14">Version 4</h2><p>Kept dimension and alpha at 16 and clip skip at 2, Added a lot more images to the dataset of particular items such as the tricorn hat, cutlass, bicorn hat and slouch hat.  The hope that was that more detailed data on these features would lead to better inferences. I'm not sure that was the case. On this version the learning rate was a lot lower than earlier versions.  I don't see a huge difference. One thing I noticed is that the items that I added images for are rendered in the inferences in the scale that they appeared in the data images, not at the scale that would mesh with the rest of the scene. Perhaps I should try making them relatively smaller in the training set so that they are approximately the same size as they will appear in the finished inference. </p>

Pirates!

p1r4t3s_v2.safetensors

Files with SHA256: b8d2c9cd8ab415e7d5b0ac7e954d7d9b3897a79510c4998a1d0239b2d85c2b70