Z-Image Base / Turbo INT8 优化版

这是我制作的 Z-Image Base 和 Z-Image Turbo 混合精度 INT8 量化版本，可直接在新版 ComfyUI 中使用。

量化时保留了敏感模块及部分关键首尾层，以减少毛边、噪点和细节损失，同时降低显存占用并提高生成速度。

在 1024 × 1024 分辨率下个人实测：

• BF16：约 1.66s/步

• 普通 INT8（不额外保留层）：约 0.85s/步

• 混合 INT8（保留关键层）：约 1.00s/步

混合版本比普通 INT8 稍慢，但画质更接近 BF16，是速度和质量之间更均衡的选择。实际性能会因显卡和工作流而异。

Z-Image Base / Turbo INT8 Optimized

These are mixed-precision INT8 quantized versions of Z-Image Base and Z-Image Turbo, compatible with the latest ComfyUI.

Sensitive modules and several key early/late layers are kept at higher precision to reduce jagged edges, noise, and detail loss while improving speed and reducing VRAM usage.

Personal benchmark at 1024 × 1024:

• BF16: 1.66s/step

• Standard INT8 without extra preserved layers: 0.85s/step

• Mixed INT8 with key layers preserved: 1.00s/step

The mixed INT8 version is slightly slower than standard INT8, but provides quality closer to BF16, offering a better balance between speed and image quality. Performance may vary depending on your GPU and workflow.

A little support goes a long way! If you’d like to help me keep creating, you can do so at https://ko-fi.com/xabsurd

• 获得更多未公开内容，和提前得至少1个月到更高质量lora输出

• Get more undisclosed content, and get at least 1 month ahead of time to higher quality lora output