MOSS-Image Base50 / SFT50 — QKV Fused（BF16 / INT8 ConvRot）

中文

这是基于 OpenMOSS 官方 MOSS-Image-Base 和 MOSS-Image-SFT 制作的等权重融合模型：Transformer 权重各占 50%。本模型是社区制作的衍生版本，并非 OpenMOSS 官方原版。

版本：BF16 版采用 QKV 融合格式，以 BF16 为主，保留 F32 辅助张量；42 层的 Q、K、V 权重按顺序拼接，转换不涉及额外的权重数值计算。INT8 版由该 BF16 底模通过 INT8 ConvRot 量化得到，部分张量仍保留浮点格式。两个版本均未额外训练，也没有烘焙 LoRA。量化可能改变输出效果。

使用：文件仅包含扩散模型权重，需要另行准备 MOSS-Image 对应的文本编码器和 VAE。BF16 版可放入 ComfyUI 的 models/diffusion_models 目录，通过 Load Diffusion Model / UNETLoader 加载。INT8 版需要支持其 ConvRot 量化格式的环境；本次尚未完成该文件的独立加载或多步出图验证。

验证：BF16 QKV 版在制作时的 ComfyUI 环境中，QKV 融合和分离格式的 LoRA 均通过加载检查，未出现未加载警告。256×256、1 步 GPU 同参数对照中，转换前后的输出 latent 完全一致；尚未完成 QKV 版的完整多步画质对照。此项验证不代表 INT8 版已经过相同测试。

封面示例的生成记录使用本 BF16 QKV 融合模型，所记录的 LoRA 均为关闭。图像元数据配置：1152×1536、Euler、50 步、CFG 3、simple，第二阶段关闭。该示例不能作为 INT8 版的出图验证。

English

This community-made derivative is an equal-weight merge of the official OpenMOSS MOSS-Image-Base and MOSS-Image-SFT Transformer weights (50% each). It is not an official OpenMOSS release.

Versions: The BF16 version uses a fused QKV layout, is primarily BF16 and retains an F32 auxiliary tensor. The Q, K and V weights in all 42 layers were concatenated in order without additional arithmetic on their values. The INT8 version was produced by applying INT8 ConvRot quantization to the BF16 model; some tensors remain in floating-point formats. Neither version received additional training or baked-in LoRAs. Quantization may change the output.

Usage: Both files contain diffusion model weights only; a compatible MOSS-Image text encoder and VAE are required separately. Place the BF16 file in ComfyUI's models/diffusion_models folder and load it with Load Diffusion Model / UNETLoader. The INT8 file requires an environment that supports its ConvRot quantization format; its independent loading and multi-step image generation have not been verified in this release process.

Validation: In the ComfyUI environment used during conversion, the BF16 QKV model loaded LoRAs with fused and separate QKV layouts without unloaded-key warnings. A matched 256×256, one-step GPU comparison produced identical output latents before and after QKV conversion. A full multi-step image-quality comparison of the QKV version has not been completed. These checks do not establish the same results for the INT8 version.

The cover example’s generation metadata identifies this BF16 QKV model, with all listed LoRAs disabled. Its recorded settings are 1152×1536, Euler, 50 configured steps, CFG 3, simple scheduler, and second stage disabled. This example does not verify INT8 image generation.

Sources and license / 来源与许可

MOSS-Image-Base: https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base

MOSS-Image-SFT: https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT

Apache License 2.0: https://github.com/OpenMOSS/MOSS-Image/blob/main/LICENSE

Credit and thanks to the OpenMOSS team for the original models. / 感谢 OpenMOSS 团队开源原始模型。