<h1 id="about-hypermoss-lush-turbo">About HyperMoss Lush Turbo</h1>HyperMoss Lush Turbo is a fast image generation checkpoint based on <a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT">MOSS-Image-SFT by OpenMOSS</a>, enhanced with a DMD Turbo LoRA specifically developed for the SFT model.It leans toward photorealistic image generation while preserving the versatility of MOSS-Image, producing high-quality results across photography, anime, illustrations, and a wide variety of artistic styles.The model retains strong prompt adherence, detailed image composition, and the ability to generate readable text within images, even at very low step counts.HyperMoss Lush Turbo is designed for fast generation, requiring only 8 steps at CFG 1, and is available in INT8 ConvRot format.<h2 id="key-features">Key Features</h2><ul><li>Strong photorealistic tendency without sacrificing artistic versatility</li><li>High prompt adherence and natural-language understanding</li><li>Supports photorealism, anime, illustration, and mixed visual styles</li><li>Good text rendering within generated images</li><li>Detailed subjects, environments, lighting, and materials</li><li>Flexible resolution and aspect ratio support</li><li>Supports natural-language and TAG-style prompting</li><li>Fast generation with only 8 steps</li><li>Available in INT8 and INT4 (w4a8) ConvRot format </li></ul><h2 id="recommended-settings">Recommended Settings</h2>Sampler: EulerScheduler: MOSS or SimpleSteps: 8CFG: 1.0HyperMoss Lush Turbo follows natural-language instructions well, making complex prompt prefixes and extensive quality tags generally unnecessary.<h2 id="prompting">Prompting</h2>HyperMoss Lush Turbo supports both natural-language prompts and Danbooru-style TAG sequences, following the capabilities of the original MOSS-Image architecture.Natural language is the recommended approach, particularly for describing subjects, expressions, poses, environments, lighting, and complex compositions.TAG-style prompting is also supported, although the model tends to perform best with clear, descriptive sentences.For best results, describe what you want to see directly. The model responds well to detailed instructions without requiring elaborate prompt engineering.Its versatility allows it to handle realistic photography, traditional anime illustrations, stylized artwork, and compositions that combine different visual styles.

HyperMoss Lush Turbo

INT4 (w4a8) version - smaller, better for old GPUs

Files with SHA256: fa836c4c84dc89c41ecb7cb0d82f360886d6586d1a2f1f7027114a668ad761a5