Currently working on dataset captioning to improve the model and expanding to other checkpoints. Soon™<h3 id="zimage-turbo-+-krea2">Zimage Turbo + Krea2</h3>Krea2 might need some more cooking, but this was the first go at it.Requires more context than MOSS.Smaller data set with more detailed captions. Trained to see how effective captioning is going (still needs work), will continue to refine and expand the data set so that it can be used for more models.Removed facial features from dataset.Trigger word: mannequinadditional trained words:<pre><code>white, transparent glass, crystal glass, (physical attributes, some poses, lighting)</code></pre><h3 id="moss">MOSS</h3> Trigger word: mannequinadditional trained words:<pre><code>white, glass, crystal glass, smooth face, retaining her facial features, eyes</code></pre>Use a descriptor in front of "mannequin" (ex: white, black, glass, crystal glass, wood, stone, gold). I did try to allow keeping or no facial features by adding "retaining her facial features" or "smooth face". If you add glasses or describe a facial feature it will most likely add them in regardless.LoRA trained exclusively on female mannequins. Images were mainly of mannequins consisting of white plastic, transparent glass, and a more reflective glass (crystal like).Decreasing weight can reduce reflectiveness of some material.It can do multiple mannequins depending on settings, anatomy may become wonky.It can generate males at lower weights, not as well or consistently.First LoRa. It is probably to broad of a data set and may be hard to get the desired generation, but it still does pretty darn well even with materials it was not trained on. I have not been able to test image to image as I do not have a workflow to use.

Female Mannequin

I believe I used the exact same data set from Zimage, maybe a little more.

Files with SHA256: 83201468e73e7aae4bd8066ece7762ae65fe17d7dd1ef86c8728ae962fd54313