<h1 id="ultra-realist-style">Ultra-Realist Style</h1>This lora is an aesthetic style blending the characteristics of gpt-image-1 and Nano Banana Pro. I particularly like the realistic skin detail, but digital illustration style both can achieve with the right prompting, and wanted to bring this local. Staring this journey I am using Z-Image-Turbo. I wanted to see how trainable it was, on top of what sort of high frequency details it was capable of (a lot of early images, and vanilla generations are very compressed looking.)<h2 id="trigger">Trigger</h2>The dataset uses a purely style-based tagging system using mainly photographic terminology besides the trigger prefix <code>ultra-realist style, ...</code> You can caption for realism as you normally would with or without triggers. <ul><li>Realism: <code>ultra-realist style, </code></li><li>Anime Realism: <code>ultra-realist style, hyper-realanime aesthetic, </code> <code>ultra-realist style, anime style, </code> </li><li>Environment: <code>hyper-realist style, ultra-detailed cinematic environments, </code> </li></ul><h2 id="weighting">Weighting</h2>The lora works well at low, and higher weights. A good threshold from testing is <code>0.5</code> - <code>1.2</code> where higher is a stronger digital illustration effect. <h2 id="dataset-info">Dataset Info</h2>The Ultra-Realist beta lora is based on a SFW dataset. Suggestive themes/gore/blood was omitted from synthetic generations. That being said, it is just a style lora, and applies to whatever context you're prompting for. It just may not have the best results in the aforementioned.

Ultra-Realist Style (Digital/CGI Illustrations)

This version of GemFX is much richer, as it uses full captioning, acting as a mini-model extender rather than just a style to existing concepts. Though because of such, and it's small sample size (actually very large for a lora) it can struggle with certain concepts. I'm hoping that will improve with more data, and eventually, just moving to finetuning rather than LoRA.Lower weighting when concept adherence struggles. Changelog:<ul><li>500% larger dataset</li><li>Supersampled dataset (clean up the noise from Nano Banana Pro's upscaler)</li><li>Full data captioning (not aesthetic captioning)</li></ul>

Files with SHA256: 9daa21a1b79f7d8b283ff4e997d288a37ddd7d843bf60d9d981ed9c133076783