<h3 id="foreword-ylvn9jxio">Foreword</h3><p>This model is the result of what I learned how LoRA models can be trained. This is still preliminary driven by the fact that there are a lot of parameters and settings which can be changed.</p><h3 id="what-the-model-does-euat8qs13">What the Model Does</h3><p>Creates a glass ball with or without content.</p><h3 id="prompt-74vw8dqvr">Prompt</h3><p>A good result can be achieved, when the topic comprises three parts like</p><pre><code>blue ocean, island in a glass ball, blue cloudy sky</code></pre><p>First and last part can sometimes be exchanged. One has to try this out. And one should experiment with something like this</p><pre><code>island in a glass ball</code></pre><pre><code>glass ball with a island  </code></pre><p>Exchanging the words in main topic can help to get what one wants.</p><h3 id="trigger-word-j7pfnhgte">Trigger Word</h3><p>I used for the training a Trigger Word. But so far I tested this Trigger word is not necessary. The Trigger Word is</p><pre><code>GlassBallConcept</code></pre><h3 id="weights-gaho8xill">Weights</h3><p>A weighting under a value over 0.9 should be avoided. Using a weight of 1.0 one gets something that looks similar to the images with one blank glass ball without background I used for the training. I trained the LoRA model using images without background. A weight of 1.0 and above leads to deformed, distorted images with undefinable background.</p><h3 id="limitations-zr28xu23u">Limitations</h3><p>It is not all the time possible to to realize a desired theme.</p><h3 id="training-is1bdze9v">Training</h3><p>I used a limited set of 9 images for the training. A captions file was also used. By use of the possibilities of AUTOMATIC1111 9 images were resulting in at least 144 images for training. This was done to prevent the so-called overfitting.</p><p>The glass balls for the training were gray, light gray, light blue, white and pure white. The background was black, light gray or white.</p><h3 id="to-do-v91mu8u3a">To-Do</h3><p>Improvement of this documentation. Refinement of the model. Test wise image creations.</p><h3 id="finally-nlwtxr51n">Finally</h3><p>Have a nice day. Have fun with the model. Be inspired!</p>

Files with SHA256: aac5e201c54996d8952047835af3b6c500c434e8ce97e64aeaa7b768c161681f