<h1 id="heading-2120">Inclined Pixels</h1><hr /><p>Pretty simple dreambooth meant to make generations into an inclined pixel art. Pixels are not perfect, but it works quite well.</p><p>Loras are working with this model, and are working quite well, even some of my heaviest ones are transferring very nicely.</p><p>Issue of edge pixels exist(most prominent on top edge), but, eh. I have an idea of how to set up a process to fix that in dataset, so, maybe in future versions, if model will be interesting to people.</p><p>NSFW works, but it was not trained to do anything in that regard.</p><hr /><h3 id="heading-2121">Training info i guess</h3><p>Simple 831 image dreambooth with tuned parameters.<br />Decoupled Learning Rate, decoupled Weight Decay, tenc clipping, etc.<br />Tenc stopped at 75%.<br />Initial offset noise is 0, noise travel is 0.0314.<br />Model config was unfrozen, token limit was set to 150, but dataset is manually tagged, and i don't recall such long strings there.<br />100 epochs, 83100 steps. Possibly could be tuned at higher res for 10-20 more epochs.</p><p></p><hr /><p>Dataset used belongs to me and was not scraped from the internet.</p>

Files with SHA256: c975db852e502c5d75d78ff71c72aadf672fab3db7fdd5b342063565775d8e4e