Ahoy!After getting inspired to train something a bit out of the ordinary again, and after trying out the Elite/Sangheili lora from Zoochman, I decided to give a non-human a shot, and what better non-human to train than the silly little grunts from halo? Pretty small but varied dataset, looks like it gave good results. Once again, tagging is key. This is V2, I allowed the WD Tagger to do its thing with V1 and once again it turned terrible with a lot of bleed. I highly advise anyone to take the time and manually tag or heavily prune WD Tagger outputs, it's worth it. Armor, backpack and gas mask are removable.<ul><li>Use realistic, photo in POSTIVE prompt to get a digital painting look. Avoid 3d tag in general as it creates really ugly results outside of photorealistic models, they look like bad 3d renders - Depends on the model though, some pony derivatives give great results.</li><li>As a more general tip, use score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, (the full string) for details. If you a put realistic, photo in NEGATIVE prompt it will give you a more anime look, and if you put in score_9, or score_4_up, alone, you will get a more flatcolor sort of look. Don't be afraid to play around with it.</li><li>Despite tagging it in every image, the weapon might persist as it bled a little bit into the concept (no idea why, but I noticed other people have had that problem), so don't forget to put weapon in NEGATIVE to get rid of it if it becomes a problem.</li><li>You can take the armor off from the grunt by not prompting (color) armor, backpack, gas mask, but the results may vary.</li><li>Note! The tag is different in the training data, it has the testing epoch number behind it. Edit it if you want to replicate my results.</li></ul>I never thought thick grunt thighs would get me acting unwise... Enjoy and post your favorite results! <pre><code>General use prompt: score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, halogrunt, 1boy, lizard, alien, yellow armor, backpack, full body, gas mask</code></pre>

Files with SHA256: 51b38365bfad456b6fd9297f7bb71dff78a9f6ca16e71dd68aa3b538151ad344