<a rel="ugc" href="https://www.youtube.com/watch?v=qEyKSFnJLuc">Troopers! We are green, and very, very mean!</a> Caution: Your model's coherence and prompt recognition -- especially with military themes -- will affect your results noticeably. Models that can do firearms above-average will give better guns, and some models will resist generating visors, helmets, or even ODSTs at all (variable with prompt). Most of the photographic examples were generated with Juggernaut Final, which performed mostly stably with this LoRA. This LoRA is still WIP. It kind of sucks. It's partially overfit, partially underfit. And it generates watermarks more than I'd like, though using <code>watermark, signature</code> in negative prompt helps reduce it. This LoRA was an absolute pain in the ass to train, if I'm being honest. Nonetheless, I hope to improve on it with dataset tweaks and retraining, then hopefully moving it to SDXL some time, too.<h3 id="heading-17"> Other tips</h3> Some annoyances can be fixed or reduced with prompting. Certain prompts just seem to kick you out of ODST-zone, which can usually be solved by describing things differently. The tendency to give bare fingers can be reduced by prompting for gloves, and <code>bare arms</code> in the negative prompt. Additionally, when you don't want a weapon, try putting <code>rifle, SMG</code> in the negative prompt if you find your ODSTs are holding things they shouldn't.

Files with SHA256: 30fda0544823af82af782873c84ebfd1d19da26b2daa251fffe8c9c806a12aa5