This custom MOSS-Image model pushes the architecture hard toward photorealism, focusing specifically on detailed skin textures and micro-details that the base model tends to smooth over. The model integrates the 8-step turbo LoRA dialed in at a gradient, so you can preserve CFG scaling and negative prompt responsiveness, while allowing the model to converge within 6 to 10 total steps. I'd recommend using the er_sde sampler and beta_57 scheduler.

While underlying blur has been significantly reined in compared to the raw checkpoint, running a light post-upscaler at the tail end polishes off any residual softness.

The model is great when you want granular control over photographic style. It's sensitive to the quality prompting, so you can go from raw, imperfect smartphone snapshots to crisp, high-res professional studio work depending on how you structure the prompt. Because of MOSS's baseline traits, when a subject gets too distant, fine details like the eyes can start to degrade.