<h1 id="dirty-moss-or-photographic-realism">Dirty MOSS | Photographic Realism</h1>A different approach to realism.Dirty MOSS is my first experimental take on the new MOSS Image model, exploring its potential for natural, believable photographic generation.Rather than chasing overly polished, cinematic aesthetics, Dirty MOSS focuses on something more subtle: making ordinary moments feel genuinely photographed.Its strongest results emerge in everyday situations: casual portraits, spontaneous snapshots, imperfect lighting, lived-in environments, and scenes that feel captured rather than carefully staged.The goal is simple: images that make you wonder whether they were actually photographed.<h3 id="two-versions-included">Two versions included</h3><ul><li>BF16: Full-precision version, preserving the original model's numerical fidelity.</li><li>CONVROT8: Optimized INT8 quantization, significantly reducing memory requirements while maintaining excellent visual quality.</li></ul><h3 id="what-to-expect">What to expect</h3><ul><li>Natural photographic aesthetics and believable everyday scenes.</li><li>Realistic skin textures, expressions, and environmental details.</li><li>Casual compositions with an authentic, unstaged feeling.</li><li>Strong performance in portraits, lifestyle photography, and ordinary environments.</li><li>A distinctive balance between realism and subtle photographic imperfections.</li></ul><h3 id="an-early-experiment">An early experiment</h3>Dirty MOSS is a first exploration of a completely new model architecture, not a finished or extensively refined ecosystem.There is still plenty to discover about MOSS, its strengths, limitations, and potential for future improvements.For now, this release is about exploring what the model already does particularly well: turning the ordinary into something remarkably believable.Experiment, explore, and share your results.<ul><li>Width and height: 896*1152 for testing or 1152*1440 for final images</li><li>CFG Scale: 2</li><li>Steps: 25</li><li>The sampling method i use is ER SDE-Simple, but you can try a few more, such as Euler-Beta , you can also experiment</li><li>Parameters for Hires.fix Lanczos, 8 steps, denoise 0.15, upscale by 1.5 CFG 2</li></ul>Sateluco

Files with SHA256: c67a8edc95c709d00b1da68b98d376372923dd6a1c1bfbab1dee1baaf78de99e