Kirei Kisegawa from Shinobi no Ittoki Update v1b (30th May 2023):<ul><li>Made minor adjustments to the dataset.</li><li>Fixed the training parameters.</li><li>Set the Noise Offset parameter to 0.05. Based on my limited experience, noise offset can help prevent a black outfit from becoming white, or vice versa. However, setting it to higher, e.g. 0.1, seemed to mess up the learning rate of D-Adaptation v3.</li><li>Retrained the model with D-Adaptation v3. (v3 is really interesting. I noticed that the learning rate occasionally increases on its own, even with Constant schedule, as observed in TensorBoard. In my latest attempt I decided to try Cosine schedule with NO warmups and NO restarts.)</li><li>Experimented with finetuning (<a target="_blank" rel="ugc" href="https://github.com/kohya-ss/sd-scripts/pull/491">https://github.com/kohya-ss/sd-scripts/pull/491</a>). Not sure if it helps or not tho.</li></ul> Prompts (v1b):-<ul><li>Trigger Words: kirei_olis, long hair, blonde hair, side ponytail, green eyes</li><li>School Uniform: aqua scrunchie, black shirt, red sailor collar, neck ribbon, long sleeves, grey skirt</li><li>Outfit 2: helmet, bodysuit</li><li>Outfit 3: orange scrunchie, blue hoodie</li><li>Outfit 4: orange scrunchie, white shirt, sleeves rolled up, blue shorts</li><li>Outfit 5: aqua scrunchie, turtleneck sweater, brown coat, brown shorts, black pantyhose</li><li>Extra: 1girl, solo (please refer to my sample images)</li></ul>Weight: 1 (you may experiment yourself) All my models were trained with 512x512 pixels. For optimal results, please upscale your generated outputs with your preferred methods. My sixth model.For some reasons, I trained a model for a character that I wasn't really interested in, again. Btw, I tried using Lama Cleaner to clean up some of my dataset this time.Hopefully some of you will like it.

Files with SHA256: 64ea8f428e544560486dfa6f5604648ab9e319a1a0232e03db74e87ee7155c4f