<h1 id="sweet-lolita-now-for-flux!-rf1vkdhi0">Sweet Lolita - Now for Flux!</h1>Version 1 is the same dataset that was used for Version 5.1 of my Sweet Lolita XL DoRA. This model is not a DoRA however, DoRA capability has yet to be added to any of the current generation programs that support Flux.<h3 id="notes:-l8uh116h2">Notes:</h3><ul><li>Quality of generation varies wildly between quants. This is new to Flux (or at least this generation of inference models, SD3 has a similar situation IIRC)</li><li>You will have to experiment with lora weight. I've found that the lora actually works better at lower weights, below 1. Most of the sample images used 0.8. As you go above 1, the style actually decreases, interestingly enough, before it eventually starts to bake. If combined with multiple other loras you may have to play around with the weights more.</li><li>General reminder about Flux prompting: while SD1.5 and SDXL single-token prompts still work (1girl,dress,photo,etc.), the entire point of the large and VRAM-hungry LLM that is used with Flux is greatly-enhanced text inference. This means that often better detail and cohesion can be found with regular-language prompting, including whole sentences. Be cognizant of generation program syntax, as you might still need to avoid using periods, for example, or line breaks. Best practices for Flux prompting are still being hotly debated, so YMMV and please let us know what works best for you! Sharing is caring :)</li></ul>On that last note... please share your generations! I love seeing what you guys come up with, and it helps me to identify flaws in the training to improve upon for future versions.<h3 id="flux-lora-training-is-&quot;expensive&quot;:-xiosg4yha">Flux LoRA training is "expensive":</h3>Please consider tipping, donating, or at least liking and posting images to help me defray the Buzz costs of using the CivitAI trainer. I'm unfortunately unable to train Flux models on my own PC right now, as Flux training currently requires significantly more VRAM than SDXL training even, and my RTX 3070 8GB GPU is slow enough at inference. I know that the training platforms are working on increased speed and low-VRAM tolerance, but at the moment my only real option is the CivitAI trainer. Your support means I can keep training, improving, and inspiring!<h3 id="note-on-use-of-the-word-&quot;lolita&quot;-on-civitai:-k8a2d2n14">Note on use of the word "lolita" on CivitAI:</h3>Civitai pretty much bans the "lolita" tag on their website. When uploading images with this LoRA in them (or even just in general), try to leave out the "lolita" tag. If it auto-tags and doesn't let you remove it, I've found that upscaling with "highly detailed" or alternatively scrubbing the EXIF data works. Lolita fashion is NOT CP! It is a legitimate J-fashion, and most lolita clothing is far too expensive for children to afford anyways! An average high-quality brand-new Lolita dress can cost anywhere from $150-500 USD!I'm doing this for fun and my own use mainly, but if you enjoy this LoRA, feel free to drop some coin my way:<a target="_blank" rel="ugc" href="http://Ko-fi.com/mysticdaedra">Ko-fi.com/mysticdaedra</a>Finally, huge thanks to the folks over on the Civitai, SD.Next, and OneTrainer discords, I couldn't have gotten into AI art and Stable Diffusion without them! You guys rock!Trained with CivitAI on-site trainer.

Sweet Lolita FLUX

Version 1:<ul><li>Initial version</li><li>To some extent the lower the weight the better</li><li>Values below 1 work best in my testing, YMMV. 0.8 is a good place to start</li><li>Due to Flux LLM, all triggers work significantly better than the SDXL versions!</li></ul>

Files with SHA256: 081b0f0621caa53e525d018c9e5a95ebc1f133dbe84fbcc93af76e3d2ed4e6f7