<h1 id="2k-ultimate-xl-(fp32)-n1zhb053h">2k Ultimate XL (FP32)</h1><h3 id="load-images-for-example-comfy-workflow-akiq68lis">Load Images for example Comfy Workflow</h3><ul><li>Built using the Full FP32 CLIP-G-Large and Improved CLIP-L vision models (The pruned CLIP-G-Large can be sourced from this model)</li><li>FP32 SDXL UNET with limited changes made (FP32 LORA's)</li><li>Works very well with SFW and NSFW Loras even those not trained at FP32</li><li>Works with SDXL Lightning 4 Step</li><li>Any 2x Upscaler can be used <a target="_blank" rel="ugc" href="https://drive.google.com/drive/folders/13kfr3qny7S2xwG9h7v95F5mkWs0OmU0D">BRSGAN 2x</a> can be found on Google Drive or the <a target="_blank" rel="ugc" href="https://huggingface.co/ai-forever/Real-ESRGAN/tree/main">2x+ model</a></li><li>For COMFY Full FP32 use --force-fp32 (For High VRAM cards)</li><li>FOR COMFY --fp32-text-enc (For Low VRAM cards)</li><li>For FORGE --clip-in-fp32 (For Low VRAM cards)</li><li>For FORGE --all-in-fp32 (For High VRAM cards)</li></ul>Note all images posted where using the FP32 model but downcast to FP16, the model still benefits from this as the source data has always been at FP32 (You can force FP32 in comfy for even higher precion but it is not required to benefit)Cover Images Remix of:<ol><li><a target="_blank" rel="ugc" href="https://civarchive.com/images/38795009">Ai_collector - Polar Bear</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/images/38699546">Blushing Female - hornystoryteller</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/images/39274507">Eye - gman_umscht</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/images/39156877">Landscape - ShallowS</a></li></ol>

Files with SHA256: 1968ae4f5834b4ccee1e7c0a469cf513a8d6ec3d78de9a7687cc111aab3d3fa0