<h3 id="must-have-tag:-76q1gtrzu">Must have tag:</h3><p>road bicycle</p><p></p><h3 id="important-optional-tags:-hm6i9pobc">Important optional tags:</h3><ul><li><p>riding bicycle</p></li><li><p>bike jersey, bike shorts, bicycle helmet, cycling shoes, cycling eyewear on head, cycling eyewear, &lt;color&gt; fingerless gloves, &lt;color&gt; gloves</p></li><li><p>front view, left side view, right side view, rear view, from behind, from above, from below, full body</p><p></p></li></ul><h3 id="non-important-optional-tags:-9hsberc6t">Non-important optional tags:</h3><p>foot on ground, foot on pedal, &lt;color&gt; handlebars, &lt;color&gt; bicycle saddle, outdoors, sports bottle, standing near bicycle</p><p></p><h3 id="as-a-bonus-it-has-learned-a-character:-9eikpr4jq">As a bonus it has learned a character:</h3><p>shiroko \(cycling\) \(blue archive\)</p><h3 id="my-thoughts-11zh4o62s"><br />My thoughts</h3><p>After so many attempts I have to conclude that without a large dataset with road bikes (most of which should be in anime style) it's impossible to have a perfectly generated road bike. The alternative would be to train a checkpoint on to understand real life road bikes, but I don't have a plan to go this far.</p>

Road Bicycle | Road Cycling

<p>I'm happy to present you the new version which now produces much better results<br />Yet, it's far from perfection. The main obstacle for me is the lack of anime/cartoon stylized images.<br />To improve the results I had to add some photos to my dataset, which affected the style but not significantly<br />After I have made 8 attempts of training this v2 model I can conclude that it's very hard for lora to understand the 3d structure of a bicycle, I suspect having 500+ images might help but currently it's not possible</p>

Files with SHA256: efdf48d5c0cb1b3a2005693b44bc45d2a3985824446f08777807e1b83dd912be