Ethical Training Data: This model was trained exclusively on original datasets. It contains no synthesized works from other authors, nor does it include any real individuals or celebrities.Please adhere to all local laws and platform guidelines when using this model. Responsible and compliant usage is mandatory.

V2 fixes the acceleration effect, but step 10 requires card drawing and may have minor limb errors;

Sampler: heunpp2 + linear_quadratic schedule
- seeds_3 + normal is better sampler.If TensorArt supports this in the future, or if it's used locally.
CFG Scale: 1 – 2.0
Steps: 8 – 20
Recommended Speed: CFG: 1, Step: 10
Recommended Quality: CFG: 1.5, Step: 20
Unless using De-distilled lora, the difference between Step 20 and Step 10 in V2 is already very small.
Prompting Tip: The more detailed and fluent your prompt, the closer the result matches your vision. Please ensure your prompts are well-structured.

V2以修复加速效果，但10步以内出图需要抽卡，会有少量肢体错误；

采样器 (Sampler): heunpp2 + linear_quadratic 调度器
- seeds_3 + normal 调度器如果以后Tensor支持、或者是在本地使用时会是最佳选择
提示词引导系数 (CFG): 1 – 2.0
步数 (Steps): 8 – 20
速度推荐：CFG：1 ，Step： 10
质量推荐：CFG：1.5 ，Step：20
V2除非使用De-蒸馏lora，Step20和Step10的图像差距已经非常小了。
提示词建议: 描述越详细、语句越流畅，生成效果越贴近您的构想。请确保提示词语法通顺、逻辑清晰。

v2 includes minor fixes and DPO color correction. For better LoRa training, v2 is the BF16 version. I've tested FP8 locally, I will only upload the FP8 version if the v2 FP8 version performs better than the BF16 version, or if someone wants to buy it.

v1 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 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 Hub or locally, if you prioritize image fidelity over generation speed, this is currently your best option.

v2 版本包含一些小的修复和 DPO 色彩校正。为了更好地进行 LoRa 训练，v2 版本采用的是 BF16 格式。我已经在本地测试 FP8 版本了，只有当 v2 的 FP8 版本比 BF16 版本效果更好，或者有人想要购买，我才会上传 FP8 版本。

v1 本检查点在女性人像、多人场景及美学构图方面表现卓越。采用多模态训练方法，它能完美适配多种艺术风格，并高质量地处理SFW（全年龄）与NSFW（成人向）任务。其核心优势在于角色多样性，能够生成涵盖全球不同族裔的人物，而不仅限于亚裔。

基于 ZIT FP8 架构构建，本模型对 TensorArt 平台高度优化，同时对低显存显卡非常友好。通过独特的机制绕开 CLIP 限制，它在处理 NSFW 内容时更加高效精准。

性能说明： 尽管恢复完整 Turbo 加速功能的尝试未获成功，但对于注重画质胜过速度的用户而言，这依然是现阶段最佳的质量首选。无论是在 Hub 在线使用还是本地部署，它都能提供顶级的生成效果。

Due to platform restrictions, I am unable to release this model as a free TSF file on TensorArt. To continue developing high-quality models like this one, your support is vital.

If you enjoy this model or plan to use it locally, please consider sponsoring me $10 on TensorArt. Your contribution directly fuels the creation of better models in the future. Thank you for believing in my work!

Ray Z-image turbo Versatile - V4 BF16