<h1 id="flux2-klein-9b-pyro's-bj">[flux2-klein-9b] Pyro's BJ</h1>New Model, New PyroThis is an experimental LoraResearch: How few ranks can you get away with?Result: 90MB lora that understands more about anatomy than your biology teacher in 5th grade.Also, it's literally the most trainable model I've ever trained with. It's so good I'm going to train it with 100k z-image-turbo images to make my own z-image-base because Alibooba is seemingly toying with our emotions. (If you want to sponsor this project shoot me a msg on discord)<h2 id="how-to-use">How to use?</h2>You can use this model with BASE and DISTILL. You can also edit images with it amazingly well.Use BASE for better prompt adherenceUse DISTILL for images faster than you can jerk offPls try my distill workflow attached as TrainingData zip-File before telling me or the world that the distill doesn't work well. Thx.Demo Images should contain workflowIn theory the LORA understands following concepts (in practice it's a bit shaky but that's why it's experimental, you'll figure it out):General: a woman sucking a penisChipmunking or whatever this shit is called: a woman cheeking a penisDeepthroat: a woman swallowing a penisWhat could this be: a woman licking a penisIf you want her to look at you in disgust: povWill experiment further until I uncovered the mysteries of flux2-klein-9b and prompt adherence is at a point where you trust the LORA more than your mom.Discord: <a target="_blank" rel="ugc" href="https://discord.gg/PSa9EMEMCe">https://discord.gg/PSa9EMEMCe</a><h2 id="technicals:">Technicals:</h2>Framework: my diffusion-pipe fork: <a target="_blank" rel="ugc" href="https://github.com/pyros-projects/diffusion-pipe">https://github.com/pyros-projects/diffusion-pipe</a>Dataset: 300imgs split evenly over all sub-conceptsTraining:65 epochs with 512p (AdamW - batch 4 - lr 2e-4)15 epochs with 768p (Prodigy Schedulefree)<h2 id="version-history:">Version History:</h2>bj_20260120_22-22-29epoch15_comfy: Initial Release

Files with SHA256: 6cb0f5363ed04d6d6c369f9a2753c1398f7175ab4c543422bdefc89c7be9efe1