<h1 id="twistedreality-xl">TwistedReality-XL</h1><h2 id="version-4-(full-sdxl-finetune)">Version 4 (full SDXL finetune)</h2>This version is a full finetune from the base SDXL checkpoint, and not base on Cyberrealistic XL. <a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0">https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0</a>These are the hyperparameters used for the training:<ul><li>8 A40s</li><li>5000 steps</li><li>lr = 1e-4</li><li>batch size = 8</li><li>gradient accumulation = 4</li><li>total effective batch size = 8 x 8 x 4 = 256</li></ul>The dataset has been captioned with either xcomposer or qwen vlm, and compressed into comma separated prompts. I've let out the pronouns, verb, so use nouns and adjectives to describe the scene.It is not very good with hardcore NSFW stuff (this is not pony, illustrious or noobai), so you need to use a Lora for this.I might look into data balancing techniques to promote rarer concepts...<h2 id="versions-3-and-below">Versions 3 and below</h2>This model is a Dora trained model on top of the excellent CyberrealisticXL checkpoint. This model enhances the realism of the photos and force the output to be in the style of the adult movie studios Playboy and ScoreGroup. I strongly suggest using Perturbed Attention Guidance (PAG) to fix some of the anatomy problems that can come up with this model. CFG can go from 1.0 to 7.0,

TwistedReality-Xl

This version has been trained with the <a target="_blank" rel="ugc" href="https://github.com/huggingface/peft/pull/2794">FourierFT PEFT</a> technique (currently a WIP PR by me), with those parameters:<ul><li>alpha = 0.04</li><li>learning rate = 1e-4</li><li>4x RTX4090s</li><li>total batch size = 32</li><li>gradient accumulation steps = 4</li><li>Total steps = 7200</li></ul>This version should be a bit better with anatomy because of the larger effective batch size.

Files with SHA256: cbc953d6d4e4389c7acf66b5543c061ed6c90b1170e5e133149845d29dfa7e46