<p>I got tired of trying to use 1940s Germans to generate commissar art for Warhammer 40K games. So, I've invested some time and learned how to make LoRAs.<br />It's my first time making a LoRA, but I think this was somewhat of a success.</p><p>This LoRA is generally good at making the uniform part, but I wasn't able to make proper insignias or aquilas.</p><h2 id="heading-842">Version 1.0</h2><p>This was my first attempt. I really didn't know what I was doing, so I made some rookie mistakes.</p><p>First of all, the LoRA wasn't directed to learn to replace uniforms, and it trained on the general vibe of the image, which was mostly Warhammer 40k commissars. When used, it affects the whole of the image, which might not be preferable if used for more specific situations.</p><p>The second mistake was that I used a phrase to activate the LoRA made up of several words: "warhammer 40k commissar". Theoretically, this could affect more of the image and conflict with other prompts.</p><p>Still, this version provides some details I couldn't capture in version 2.0 that were ignored. Also, because of the learned vibe of Warhammer 40K art, it reproduces that while generating images.</p><h2 id="heading-2839">Version 2.0</h2><p>This time I actually included the painfully crafted captions for the training images in the learning process. Which should have helped focus LoRA training on the uniform. By extension, that should have limited the LoRA's effect on the rest of the image, making it more versatile. Also, this time I used regulation images while training, hopefully improving the end result.</p><p>The result of all of this was that the LoRA didn't learn the overall style of Warhammer 40K art (to some disappointment) and now creates more consistent commissar uniforms. Some details that were accidentally learned in version 1.0, like the horizontal golden cords on the commissar shirt or the unique red sash, are rarer. Though I learned that prompting for the sash made it more frequent.</p><h2 id="heading-2837">Version Comparison</h2><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/88db7ac8-9ba6-45e4-9b48-40d909bb0915/width=525/88db7ac8-9ba6-45e4-9b48-40d909bb0915.jpeg" /></p><p>The prompt was a simple concept: "girl, black uniform, arms crossed". Although V2 has affected more of the image than just the uniform, the impact is way less than that of V1, which didn't follow that the prompt "black uniform" was connected to the commissar.<br /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/3d1f294a-6216-4316-8377-fb20bbe4e7aa/width=525/3d1f294a-6216-4316-8377-fb20bbe4e7aa.jpeg" /></p><p>What would be perfect is to completely replace the military uniform on the right with a commissar uniform, leaving the rest of the character untouched. Alas, V2 and V1 affect the image more than just the character.</p><h2 id="heading-2838">Future plans</h2><p>I am a bit miffed that some of the details are missing in the current LoRA, and I am thinking of ways to bring that out. This will most likely involve a lot of trial and error. If I am successful, I will post a new version here.</p><p>Although I think the concept of the commissar uniform is better understood by V2 than V1, the LoRA still affects the rest of the image in a significant way. I think maybe better captioning for the training and more processed training data would result in a better version.</p><p>Negative TIs used: <a target="_blank" rel="ugc" href="https://civarchive.com/models/4629/deep-negative-v1x">Deep Negative V1.x</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/models/16993/badhandv4-animeillustdiffusion">badhandv4</a></p>

Files with SHA256: 8cfb1f82a86aa033af44b2fd6054c117b894c58d4a1848628eb60782381f7546