<h1 id="flux:-ame4ku3xn">Flux:</h1>A Flux dev version of the SDXL images and captions. Please check the version description for more information. I've included the preview wildcards as well. I tried to mirror the SDXL training workflow more closely in an attempt to homogenize the process for both models. Captions were closer to the mixed tag/natural language format of the SDXL process rather than my previous overly verbose Flux training prompts.<h2 id="overview:-ezdf1a1qt">Overview:</h2><ul><li>Model consists of five of her cosplay sets, Alya, Kotori, Momo, Nagatoro, and Rin.</li><li>Masked training was not used.</li><li>Regularization was not used.</li><li>Dreambooth trained using the <code>sd3-flux.1</code> branch of the Kohya_ss scripts.</li><li>Dreambooth trained on Flux.1-Dev (fp16), then LoRA extracted.</li></ul>The training was tested using the MGAS comparison method to find the best extracted-epoch. The introduction article can be found <a target="_blank" rel="ugc" href="https://civarchive.com/articles/6327/an-introduction-to-the-mgas-comparison-method-and-scripts">here</a>.<h2 id="usage:-a360bxuie">Usage:</h2><ol><li>Intended for use with Flux.1-Dev. Performs OK on other realistic focused models.</li><li>Size: 1024,1024, Steps: 20, Sampler: Euler-Simple, CFG/Guidance scale: 1.0/3.0-3.5, Strength: 1.0-1.1</li><li>[Optional] Adetailer. Was NOT used for preview images.</li><li>All cosplays are tied together by the common tag, {cata7pln}, which consists of two rare tokens on either side of a number.</li><li>Add the following tags in your prompts for the individual cosplays:<ul><li>Alya - {Alya cosplay, long pastel pink hair, blue eyes, red ribbon in hair, red bowtie with white stripes, grey blazer, blue vest and skirt, white undershirt, white thighhighs}</li><li>Kotori - {Kotori cosplay, long red hair, twintails, red eyes, two-tone bow in hair, black tie, red blazer, white undershirt, red skirt, black thighhighs}</li><li>Momo - {Momo cosplay, short brown hair, red eyes, large green earrings, black choker, white shirt collar, pink sweater, red bowtie, blue skirt, white leg warmers}</li><li>Nagatoro- {Nagatoro cosplay, long black hair, hazel eyes, hair clips in hair, white button-up shirt, rolled up sleeves, black pleated skirt}</li><li>Rin- {Rin cosplay, long brown hair, twintails, green eyes, black bow in hair, red turtleneck sweater, white cross logo, black pleated skirt, black thighhighs}</li></ul></li></ol><h2 id="future-outlook:-ouylihdc0">Future Outlook:</h2>I plan to further refine my training workflow using my favorite cosplayers as I continue to build the dataset for the eventual nobody checkpoint.==========================================================<h1 id="sdxl:-icifdozpw">SDXL:</h1>Testing out my training workflow on another fantastic cosplayer.<h2 id="overview:-jhrbbomyd">Overview:</h2><ul><li>Model consists of five of her cosplay sets, Alya, Kotori, Momo, Nagatoro, and Rin.</li><li>Masked training was used for more flexible backgrounds.</li><li>Regularization set was captioned with GPT-4o for greater control of details.</li><li>Original Dreambooth trained using Constant AdaFactor with Min_SNR_Gamma loss.</li></ul>It was extensively tested using the MGAS comparison method (<a target="_blank" rel="ugc" href="https://civarchive.com/articles/6327/an-introduction-to-the-mgas-comparison-method-and-scripts">here</a>) during both training and image generation, while an example regularization prompt can be found in the VDC article (<a target="_blank" rel="ugc" href="https://civarchive.com/articles/6264/a-simple-script-for-visualizing-captions-in-large-datasets">here</a>).<h2 id="usage:-9ivine1y2">Usage:</h2><ol><li>Trained on, and works best with, RealVisXL V5.0. Works well with other popular realistic XL checkpoints such as JuggernautXL, jibMixRealisticXL, mklanRealisticXL, copaxTimelessXL, etc.</li><li>Size: 1024,1024, Steps: 50, Sampler: DPM++2m Karras, CFG scale: 7, Strength: 1.0<ul><li>DPM++SDE/Euler-a Karras at 30 steps works as well.</li><li>832x1216 and 1216x832 (i.e. 2:3 &amp; 3:2) tend to work well too.</li></ul></li><li>[Optional] Adetailer, with denoise around 0.4.</li><li>All cosplays are tied together by the common tag, {cata7pln}, which consists of two rare tokens on either side of a number.</li><li>Add the following tags in your prompts (see preview images for examples) for the individual cosplays:<ul><li>Alya - {Alya cosplay, long pastel pink hair, blue eyes, red ribbon in hair, striped red bowtie, grey blazer, blue vest and skirt, white undershirt, white thighhighs}</li><li>Kotori - {Kotori cosplay, long red hair, twintails, red eyes, two-tone bow, black tie, red blazer, white undershirt, red skirt, black thighhighs}</li><li>Momo - {Momo cosplay, short brown hair, red eyes, large green earrings, black choker, white shirt collar, pink sweater (or pink windbreaker), red bowtie, blue skirt, white leg warmers}</li><li>Nagatoro- {Nagatoro cosplay, long black hair, hazel eyes, hair clips, white button-up shirt, rolled up sleeves, black pleated skirt}</li><li>Rin- {Rin cosplay, long brown hair, twintails, green eyes, black bow, red turtleneck sweater, white cross logo, black pleated skirt, black thighhighs}</li></ul></li></ol>==========================================================

Caticornplay [Flux/XL] (r557)

Initial Release - Version 1.0Dreambooth trained on 75 images, broken up into 5 sets of 15 images with one set for each cosplay (one concept), using the <code>sd3-flux.1</code> branch of the Kohya_ss scripts. Training was done for 175 epochs in batches of 1. Finally, the LoRA was extracted using Koyha_ss script's utility tab.<ul><li>Total trained steps: (5 x 15) x 1 x 175/ 1 = 13,125- or 2,625 per cosplay/concept</li><li>Extracted LoRA rank/alpha = 32/(dynamic)</li></ul>

Files with SHA256: 8b427de09e240e23c3727cfb6c05d65c2c9da84660522a776b7686e307659918