About this version

Sampler: heunpp2 + linear_quadratic schedule

seeds_3 + normal is better sampler.

CFG Scale: 1 – 2，Steps: 8 – 20

Recommended Speed: CFG: 1, Step: 10

Recommended Quality: CFG: 1.5, Step: 20

Prompting Tip: The more detailed and fluent your prompt, the closer the result matches your vision. Please ensure your prompts are well-structured.

v1.2 this checkpoint excels in generating female portraits, group scenes, and aesthetic compositions. Utilizing a multimodal training approach, it adapts seamlessly to diverse art styles and handles both SFW and little NSFW tasks with high fidelity. A key feature is its character diversity, ensuring a wide range of ethnicities beyond just Asian characters.

Built on the ZIT FP8 architecture, it is highly optimized for TensorArt and remains friendly to low-VRAM GPUs. It employs a specialized mechanism to bypass CLIP restrictions, significantly enhancing its capability in handling NSFW content.

Performance Note: While attempts to fully restore Turbo acceleration were unsuccessful, this checkpoint stands as the premier choice for quality-focused users. Whether on the TensorHub or locally, if you prioritize image fidelity over generation speed, this is currently your best option.