<h3 id="preface-330nxncet"><strong><span style="color:rgb(76, 110, 245)">Preface</span></strong></h3><p>When I started with AI image generation I tried to add a pyramid to a field of roses. That was much more complicated than expected. Most of the time the pyramid looked like a real egyptian one with a rough surface and a smooth or other type of tip.</p><p>The new LoRA I created works after one creation try out of the box with quite good results. My magic wand LoRA is more complicated to handle than the pyramid.</p><p>The reason for the last statement is simple. Everybody inclusive AI knows how the sizes of a pyramid are in the real world normally. It makes no difference if it is an Egyptian pyramid or a Maya pyramid.</p><h3 id="what-the-lora-does-xtkgttvat"><strong><span style="color:rgb(76, 110, 245)">What the LoRA Does</span></strong></h3><p>Creates a black pyramid or modifies a pyramid. The modification should be a black pyramid with shiny sleek sides.</p><h3 id="training-images-qv9fjmc62"><strong><span style="color:rgb(76, 110, 245)">Training Images</span></strong></h3><p>For the first attempt of the LoRA creation, I have prepared a set of 15 training images. All images look like the black pyramid in Figure 1. The background is slightly different and the size of the pyramid changes slightly.</p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d25d336f-bafd-4c95-b321-70c8108cbef9/width=525/d25d336f-bafd-4c95-b321-70c8108cbef9.jpeg" /><em>Figure 1: Black pyramid with blurry background</em></p><p>All images have a size of 512 x 512 pixel. In sum I used 320 images for the training. Mirroring and buckets of different size results in the number of total training images.</p><h3 id="self-criticism-hkfhvqsxg"><strong><span style="color:#4c6ef5">Self-criticism</span></strong></h3><p>After working with the new LoRA I took a look at the trainings images and figured out, that some of the images are not the way, they should be.</p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c65c3f9d-a6c0-45fc-993f-d3de5dc48e36/width=525/c65c3f9d-a6c0-45fc-993f-d3de5dc48e36.jpeg" /><em>Figure 2: Bad trainings image</em> </p><p>As one can see if one take a closer look, or if the image is for a review large enough, the surface is not smooth, as it should be for the realisation of my idea.</p><p>The frame or a frame around the sides is for the first versions of the LoRA model acceptable.</p><h3 id="to-do-2t4jkxqcp"><strong><span style="color:rgb(76, 110, 245)">To-Do</span></strong></h3><p>Improvement of this documentation.</p><p>Preparation of better trainings images for the next version of the LoRA model.</p><h3 id="final-words-2546ebmiv"><strong><span style="color:rgb(76, 110, 245)">Final Words</span></strong></h3><p>Have a nice day. Happy LoRA testing. Be inspired!</p>

Files with SHA256: f80c2a6533bd8a646da5b4048b31c0e8e7bcbaa5fe905e7d0d8685ef1cefd57b