<h1 id="flux:-9x547k69r">Flux:</h1>A Flux dev version of the SDXL images. Please check the version description for more info. ¡NOTE! FluxDev alone still has trouble with nsfw. Either try one of the suggested LoRA or checkpoints.<h2 id="overview:-p7kxh7l31">Overview:</h2><ul><li>Model consists of three outfits, Main, Secondary, and Nude.</li><li>Masked training was not used.</li><li>Regularization was not used.</li><li>LoRA trained using the <code>sd3-flux.1</code> branch of the Kohya_ss scripts.</li><li>LoRA trained on Flux-Dev2Pro for use with Flux.1-Dev (fp8/fp16).</li></ul>The training was tested using the MGAS comparison method to find the best 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:-dntctergc">Usage:</h2><ol><li>Intended for use with <a target="_blank" rel="ugc" href="https://civarchive.com/models/618692/flux">Flux.1-Dev</a>. For full nsfw capabilities consider other models such as <a target="_blank" rel="ugc" href="https://civarchive.com/models/833086?modelVersionId=1462430">Colossus Project Flux - v5.0</a> or <a target="_blank" rel="ugc" href="https://civarchive.com/models/118111?modelVersionId=1273458">Copax Timeless - XPlus_4.</a></li><li>Size: 1:1, 3:4, 2:3, and 9:16, Steps: 25-40, Sampler: DPM++2M-Beta, CFG/Guidance scale: 1.0/3.0-3.5, Strength: 1.0.</li><li>[Optional] Adetailer. Was NOT used for preview images (Highres Fix only).</li><li>All outfits are tied together by the common tags, {aml7pln, woman, cosplay}.</li><li>Incorporate the following tags in your descriptive FLUX prompts (see preview images for examples) for the individual outfits:<ul><li>Main - {red headband, cropped red jacket, goggles hanging around neck, white blouse, brown corset, belt, brown shorts, thighs, red and white thighhighs, high heels}</li><li>Secondary - {red headband, goggles perched atop head, white shirt, red bloomers, white stockings, slippers}</li><li>Nude - {red headband, nude, breasts, nipples, navel, stomach, thighs, pussy peek, red and white thighhighs}</li></ul></li><li>Variations &amp; Tips:<ul><li>nsfw preview images combine the LoRA, <a target="_blank" rel="ugc" href="https://civarchive.com/models/643366?modelVersionId=725939">Flux Improved Female Nudity</a> (~0.7).</li><li>Try adding long dark hair or vary the color.</li><li>Both blue and brown eyes.</li><li>Variable chest size and several poses (SFW &amp; NSFW).</li><li>Black bikini, wet clothes, see-through shirt, and heart pasties.</li><li>Close-up, cowboy, and full-body shots. Front and back views.</li></ul></li></ol>==========================================================<h2 id="sdxl:-sen9g77i3">SDXL:</h2>My first attempt at making a consistent character cosplayer. This one is solely focused on generating a realistic Amber from Genshin Impact. The used wildcards for preview images are provided as well.<h2 id="overview:-ww9h4fuur">Overview:</h2><ul><li>Model consists of three outfits, Main, Secondary, and Nude.</li><li>Masked training was used with OneTrainer for more flexible backgrounds.</li><li>Regularization was used with a large set of ground-truth images and GPT-4V captions for detail control.</li><li>Dreambooth trained (Constant Adafactor with Min_SNR_Gamma loss) and LoRA extracted for improved results.</li></ul>The training was tested using the MGAS comparison method to find the best 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:-sqsmkrd3j">Usage:</h2><ol><li>Trained on, and works best with, RealVisXL V5.0. Also works well with the big and popular realism models such as Juggernaut XL and CyberRealistic.</li><li>Size: 1:1, 3:4, 2:3, &amp; 9:16 XL, Steps: 25-35, Sampler: DPM++2M-SDE, CFG scale: 5-7, Strength: 1.0</li><li>[Optional] Adetailer can help, with denoise around 0.3-0.4. Preview images were Highres Fixed with a denoise of 0.35-0.4 and upscale of 1.5x.</li><li>All outfits are tied together by the common tags, {aml7pln, woman, cosplay}.</li><li>Add the following tags in your prompts (see preview images for examples) for the individual outfits:<ul><li>Main - {red headband, cropped red jacket, goggles around neck, white blouse, brown corset, belt, brown shorts, thighs, red and white thighhighs, high heels}</li><li>Secondary - {red headband, goggles on head, white shirt, navel, stomach, red bloomers white stockings, slippers}</li><li>Nude - {red headband, nude, breasts, nipples, navel, stomach, thighs, pussy peek, red and white thighhighs}</li></ul></li><li>Variations:<ul><li>Both blue and brown eyes.</li><li>Variable chest size and several poses (SFW &amp; NSFW).</li><li>Black bikini, wet clothes, and see-through shirt.</li><li>Close-up, cowboy, and full body shots. Front and back views.</li></ul></li></ol>==========================================================

Amber Cosplayer CC [Flux/XL] (r557)

Initial Release - Version 1.0Dreambooth trained on 60 images. Training was done for 200 epochs in batches of 4 with a low LR. This dataset was regularized by a 1:1 balanced set of 5200 ground-truth images of real women, prepared by our resident Maestro, SECourses (see his Patreon), and captioned via GPT-4V. Finally, the LoRA was extracted using Koyha_ss scripts at bf16.<ul><li>Total trained steps: (60) x 2 x 200 / 4 = 6000</li><li>Extracted LoRA Rank/Convolution: 16/8</li></ul>

Files with SHA256: a512c68d2a1cd0dbbed18e0a6b1866aa6107756528a0dd0bac6373f3872b015a