trigger <span style="background-color:rgb(24, 24, 27);font-family:Lato, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, &quot;Helvetica Neue&quot;, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;, &quot;Segoe UI Symbol&quot;, Unifont;font-size:14px">LJNERATIONX_rubberStrength around .7-.8No need to say "figure" or the like. I tried to train it to treat them like people rather than toys.If you were a kid in the 80s, you probably had some of these guys, even if you didn't like wrestling. That said, I went out of my way to train this to not spit out wrestlers, just big dumb rubber people of any ilk.Other terms it likes- scuffed up, worn paint, new in package (see samples to do the boxes- seems leaving trigger off helps with boxes only!), sealed under a plastic bubble on a blue/black card, mold spots, hand-painted, WWF logoTraining notesThis may not look it, but the data set was a beast to get trained properly. There were almost no female figures made, which was a huge challenge. As such, fellow wrestling dorks and interested trainers, I did the following for the data set:<ul><li>I had to use a few custom figures which required a lot of editing and a whole lot of retraining- Z absolutely picks up on the texture difference between real figures and hand painted customs</li><li>Yanked a few select pics of CharaPro Japanese figures and Sofubi figures, editing out the seams at articulation points</li><li>The packaging for these figures was very inconsistent with placement of text, logos, etc. As such, I spent an entire day editing them to remove price tags, severe creases, bilingual text on foreign packages, etc.</li><li>I did at least 150 training runs. I am insane.</li><li>Ultimately, what worked was training at 512, batch size 20, 400 steps, adafactor sigmoid sampler @ LR 0.001, and differential guidance cranked to 5</li></ul>

Files with SHA256: 495cd8208144a48d48d168aae4761ffa7a0f987ed216041273d7c3616351dead