I like weird, quirky portraits and have already uploaded a quite a few here on Civitai. That's how I came up with the idea of creating a LoRA for it. For the training I mainly used my own images that I created with Flux, see for example <a target="_blank" rel="ugc" href="https://civarchive.com/collections/6451635">this collection</a>. More details about the training can be found further down in the text. The result didn't turn out the way I imagined it, but sometimes interesting, cool pictures emerge. Maybe someone here has a use for it 😊 I have used the following phrases/keywords during training and these can enhance the effect:<pre><code>(neon) face paint / make-up / hair / colors
special (eyes / hair / skin)
(lots of) flowers / jewellery (in the hair)
freckles, glasses, eyes closed</code></pre>In my experience, the trigger word (w3p0rtra1t) is not absolutely necessary and can also be omitted and a LoRA strength in the range between 0.5 and 1.5 leads to interesting results. I've also created a <a target="_blank" rel="ugc" href="https://civarchive.com/collections/6951715">collection of all my showcase, test and trial images</a> I've made with this LoRA, all with public prompts and settings. Feel free to use them as a template or starting point for your images, or remix them in other ways 😊<h3 id="known-negative-side-effects-s19v23ejm"> Known negative side effects</h3>Unfortunately, when using the Lora, I found that it often produces very plastic-like faces at high strengths and depending on the keywords. Perhaps this is due to the many pictures with face paint / heavy make-up that I used for the training. Be aware of that and maybe use a lower strength or different keywords to avoid that effect.<h2 id="training-details-j0yi0wcks"> Training details</h2>I used 137 images for the training, all of which have a resolution of 512x512. Especially because I had unfortunately only created many of them in the small resolution and I had read that 512x512 also works quite well with Flux. In terms of image selection, I tried to choose a balanced and diverse mix of images (different genders, skin colours, styles, etc.) and not to overemphasise any of them. I then generated 20 epochs and selected epoch 17 in a lengthy process of elimination, as it provided the most interesting results.As this is my second LoRA and I just roughly know what I am doing, I welcome any feedback or suggestions for improvement 😄

Files with SHA256: af27388cd90551260d8a1785c00da2af1c4f2c5bf2f878bc8a840a1cfa098278