kongbai-84/AnimaInt8Quantizer-Portable

Anima INT8 Quantizer Portable

这是一个面向 Windows 的 Anima INT8 量化便携工具。它将兼容的 Anima 模型权重转换为 Forge Neo / ComfyUI 可加载的 INT8 ConvRot 权重，以降低显存和存储需求。

This is a portable Windows utility for Anima INT8 quantization. It converts compatible Anima model weights into Forge Neo / ComfyUI-compatible INT8 ConvRot checkpoints to reduce memory and storage requirements.

支持的模型 / Supported Models

• Anima 原版 2B / original 2B：支持自动识别模型结构，当前检测到 28 个 net.blocks。
• Anima 2.9B：支持自动识别模型结构，当前检测到 40 个 net.blocks。

The quantizer automatically detects the number of blocks, keeps the final output layer in the checkpoint, and produces Forge Neo / ComfyUI-compatible INT8 weights.

模型下载 / Model Downloads

• Anima 原版 2B / original 2B：请从 CircleStone Labs Anima model repository 获取有权使用的模型文件。
• Anima 2.9B：Civitai - Anima 2.9B

重要：本工具包不包含任何 Anima 模型权重。

Important: this package does not include Anima model weights.

请仅下载和使用你有权使用的模型文件，并遵守模型卡、许可证及平台条款。

Only download and use model files you are authorized to use, and follow the applicable model card, license, and platform terms.

包含内容 / Package Contents

完整便携目录包含：

• AnimaInt8Quantizer_Portable.exe
• _internal 运行时依赖目录
• AnimaInt8Quantizer.py 可审阅的量化源码
• AnimaInt8Quantizer.spec PyInstaller 构建配置
• 启动脚本：D:\Ai\工具\启动AnimaInt8量化工具.bat

The portable directory contains the Windows executable, its bundled _internal runtime dependencies, reviewable source code, and the launcher script. The package does not require a separate comfy_kitchen installation; the ConvRot INT8 implementation is built into the tool.

使用方法 / Usage

1. 完整解压便携目录，不要只复制 exe 文件。

2. 运行 AnimaInt8Quantizer_Portable.exe。

3. 选择输入的 .safetensors 模型和输出文件路径。

4. 点击开始量化，等待任务完成。

5. 将生成的权重放入对应的 ComfyUI / Forge Neo 模型目录，并按实际工作流加载。

6. Extract the complete portable directory; do not copy only the executable.

7. Run AnimaInt8Quantizer_Portable.exe.

8. Select the input .safetensors model and output file path.

9. Start quantization and wait for completion.

10. Place the generated checkpoint in the appropriate ComfyUI / Forge Neo model directory and load it in your workflow.

界面支持中文 / English 切换，并针对 Windows 高 DPI 显示器启用了 Per-Monitor DPI 感知。工具默认从第 3 个 block 开始量化到最后一个 block，也可以通过源码命令行参数调整范围。

The GUI supports Chinese / English switching and Per-Monitor DPI awareness for high-resolution Windows displays. By default, quantization starts at block 3 and continues through the final detected block; the source CLI can adjust the range.

输出格式 / Output Format

输出文件使用 Forge Neo / ComfyUI 的 int8_tensorwise 格式，并包含：

• INT8 weight
• 逐输出行 weight_scale
• comfy_quant ConvRot 配置

The output uses Forge Neo / ComfyUI's int8_tensorwise format with INT8 weights, per-output-row scales, and ConvRot metadata.

上游模型与许可证 / Upstream Models and Licenses

Anima 是 CircleStone Labs 与 Comfy Org 合作开发的文本生成图像模型，主要面向动漫、插画及其他非写实内容。

Anima is a text-to-image model developed by CircleStone Labs and Comfy Org, primarily intended for anime, illustration, and other non-photorealistic content.

本工具包是非官方的便携量化辅助工具，不代表 CircleStone Labs、Comfy Org、NVIDIA、Hugging Face 或 Civitai 的认可、授权或验证。

This package is an unofficial portable quantization utility. It is not endorsed, authorized, or validated by CircleStone Labs, Comfy Org, NVIDIA, Hugging Face, or Civitai.

Anima 模型及其衍生模型的使用、修改和分发须遵守原始模型仓库的最新许可证和模型卡要求。Anima 同时是 nvidia/Cosmos-Predict2-2B-Text2Image 的衍生模型，在适用范围内还须遵守 NVIDIA Open Model License Agreement。

Use, modification, and distribution of Anima models and derivatives must follow the latest license and model-card requirements from the original model repository. Anima is also derived from nvidia/Cosmos-Predict2-2B-Text2Image; the NVIDIA Open Model License Agreement applies where relevant.

本工具不对模型权重、量化后权重或第三方依赖授予额外权利。压缩包中的第三方软件组件分别适用其各自许可证。

This tool grants no additional rights to model weights, quantized weights, or third-party dependencies. Third-party components remain subject to their respective licenses.

模型输出图像的使用范围与模型权重不同，请以原始模型卡和许可证的最新条款为准。使用者须自行确保下载、量化、分发和使用行为符合适用法律及所有相关许可证。

The rights covering generated images may differ from those covering model weights; refer to the latest original model card and license. Users are responsible for ensuring that downloading, quantizing, distributing, and using the materials complies with applicable law and all relevant licenses.

免责声明 / Disclaimer

本工具按“现状”提供，不附带任何明示或默示保证。使用者需自行承担模型下载、量化结果、硬件兼容性、生成内容及后续传播产生的风险和责任。

This tool is provided “as is” without warranties of any kind, express or implied. Users assume all risks and responsibilities arising from model downloads, quantization results, hardware compatibility, generated content, and subsequent distribution.