still experimental and WIP: I've gotten v3.1 to the point where it works about half the time and with a variety of clothes, but I still want one that will work unanimously. I still have more to do to improve it but I think it's good enough for those with their own hardware to give it a spin. I ran a few tests, there is about a 50/50 success rate in creating the right cutout with the onsite generator. (the "failures" will look fine, just no cutout effect). If you're paying buzz to generate, I'd recommend holding back on this version unless you have lots to spare. Seems like with Flux LoRAs it is really important to be patient follow an iterative process.triggerwords: ubcutout, narrow cutout, large cutout, single cutout, separate cutouts<h2 id="training-notes-vowd74396">Training Notes</h2><h3 id="v6.0-p6sder6j5">v6.0</h3><ul><li>I'm doing some more thorough testing identifying the blocks used for top styles<ul><li>seed: 400x400 1851087289<ul><li>Major blocks: 7 9 10, 11!, 12</li><li>Minor blocks: 8 13(pose?) 22</li></ul></li></ul></li></ul><h3 id="v5.0-cnxy9rutn">v5.0</h3><ul><li>it's okay, it isn't as consistent as 3.1 but not bad, i will upload if i get a stronger version</li><li>exact same settings as 3.1 except with single block transformers 5,8,9,10</li><li>using new dataset from 4.0</li></ul><h3 id="v4.0-(failedcrashed)-li5s3niaq">v4.0 (failed/crashed)</h3><ul><li>welp it crashed 1/3 of the way in (something to do with the sample images) and i didn't have it save resume data, so i just have the first epoch to test from</li><li>used 3.1 to generate some new data, to improve clothing texture/materials<ul><li>replaced dataset images that were oversmoothed/clearly upscaled</li><li>replaced a few dataset images that have a white background</li></ul></li><li>cosine with 3 repeats, T_Max = 402??<ul><li>did i mess this up? it should be 1/3 of the training steps no?</li></ul></li><li>LR back to 0.0001</li><li>w/regularisation images</li><li>dataset 67 images, 4020 training steps total</li></ul><h3 id="v3.1-vmyt34i1o">v3.1</h3><ul><li>Same as 3.0 except with cosine LR<ul><li>(next time need to find correct config settings to get cosine with repeats)</li></ul></li></ul><h3 id="v3-1vg8hiooq">v3</h3><ul><li>Dataset adjusted to 50 images, pruned bad images and added some better ones</li><li>Simplified existing tags, also added support for <code>large cutout</code>, <code>narrow cutout</code>, <code>zipper cutout</code>, <code>single cutout</code>, <code>separate cutouts</code></li><li>Learning rate increased slightly to 1.2e-4<ul><li>reduced to 0.5e-4 at e1400 (started looking overcooked around e2000)<ul><li>this was not productive, next time i should start from closer to when it started looked burned or change the LR/schedule altogether</li><li>upon testing, e2400 is actually not bad, maybe I should have let it bake for a bit longer</li></ul></li></ul></li><li>rank/alpha decreased to 8 (~85MB), probably bigger than it needs to be for a clothing style LoRA</li><li>4000 steps</li></ul><h3 id="v2-kyrod09hc">v2</h3><ul><li>pretty much default ai-toolkit settings</li><li>I think if i trained much longer at the same settings it would work, it gets the desired effect ~10% of the time with minimal side effects<ul><li>tends to simplify fabric textures, and hard to get the fabric in sharp focus</li></ul></li></ul>

Files with SHA256: 4e4c5860d0b4503b6a35a1f941aa4fce0ee4a881d4aa322a4f42ad2bfe71cbda