Gloria Guida is one of the main interpreters of the Italian sexy comedies of the seventies, as well as other films mainly with an erotic background. All the images of Gloria that I have used come from her period of activity which goes from 1974 (when she was 19 years old) to 1979.v0.7 x30batch1 despite 35400 steps (but with Train batch size at 1) is a little undertrained. TE instead seems to me quite well trained. Will retrain this dataset on a Google Colab notebook.. I can't freeze my GPU for 5 hours.V0.6.4 was trained on a dataset of 59 images mostly from 70's magazine scans and film stills, enhanced with Photo AI, captioned with WD14 then manually edited. In some of the images used, Gloria is completely naked, but is shot from behind. However, her beauty has always been elegant and never obscene. In Italy they had given her the nickname of "queen of showers".I've recently started using D-adaptation as optimizer, which, although it gives better results, needs more tests to achieve optimal results, so expect at least one more release before v1.0 (I'm a perfectionist). In future versions also I hope to arrange a better dataset.

Gloria Guida (LoRA)

Trained on dataset of 59 Topaz photo AI enhanced images (x 9 repeats) with WD14 captioning (manually edited) and 531 regularisation images (x 1 repeats)Kohya settings:<ul><li>Optimizer: Dadaptation (D-Adaptation AdamPreprint optimizer)</li><li>Optimizer extra arguments: "decouple=True" "weight_decay=0.01" "betas=0.9,0.99"</li><li>Mixed precision: bf16</li><li>Save precision: bf16</li><li>LR Scheduler: Costant</li><li>LR warmup: disabled</li><li>Learning rate: 0.5</li><li>Text encoder learning rate: 0.5</li><li>Unet learning rate: 0.5</li><li>Network Rank (Dimension): 128</li><li>Network Alpha: 1</li><li>Max resolution: 576,576</li><li>Min SNR gamma: 5</li><li>Noise offset: 0.05 </li><li>Clip skip: 1</li><li>Train batch size: 6</li><li>Epoch: 10</li><li>Steps: 1770</li></ul>

Files with SHA256: 588e91b2d430db393b5f2621dce7c2757d5b08d14c16a699d62e862887ca3ccf