Views
No views yet
A ComfyUI custom node that quantizes diffusion models (safetensors) to NVFP4 / FP8 (e4m3fn) / GGUF in place — with a per-architecture blacklist that protects structure-sensitive layers so the quantized model actually loads and renders correctly.Auto (Universal)mode works on any architecture (T2I/I2I/T2V/I2V, including brand-new models) without per-model tuning.
binyuan Universal Quant Converter (NVFP4/FP8/GGUF),分类 binyuan/Advanced。
读取 ComfyUI 的 checkpoints / unet / diffusion_models 里的源模型,按所选格式 + 模型类型量化,输出新的 safetensors(NVFP4/FP8) 或 .gguf 文件。自动承袭原模型全部元数据,并写入 comfy_quant 标记,让 ComfyUI 正确识别量化层。pip install comfy-kitchenpip install gguf(通常随 ComfyUI-GGUF 安装)ComfyUI/custom_nodes/ComfyUI_binyuan_NVFP4_Converter/,重启 ComfyUI。binyuan/Advanced → binyuan Universal Quant Converter。input_model(下拉自动列出 checkpoints/unet/diffusion_models 里的模型)。output_format:NVFP4 / FP8 (e4m3fn) / GGUF。model_type(决定黑名单策略):
Auto (Universal):通吃任意架构,无需逐模型调参(推荐首选)。gguf_qtype:Q8_0/Q5_0/Q4_0/BF16/F16 等。save_to_source_dir=true 存到源模型同目录;或在 custom_output_dir 手填目录。填 output_filename。status 端口返回结果信息。| 参数 | 说明 |
|---|---|
| input_model | 源模型(自动扫描 checkpoints/unet/diffusion_models) |
| output_format | NVFP4 / FP8 (e4m3fn) / GGUF |
| model_type | 黑名单策略;Auto 通吃,或选具体架构 |
| gguf_qtype | 仅 GGUF 生效(Q8_0/Q5_0/Q4_0/Q6_K/Q5_K/Q4_K/Q3_K/Q2_K/IQ4_*/BF16/F16) |
| full_precision_mm | NVFP4/FP8 推理模式:开=反量化到全精度做矩阵乘(任何 GPU 能跑,最稳,不省显存);关=原生量化矩阵乘(真省显存/加速,需 Blackwell+cu130+comfy_kitchen) |
| device | cuda / cpu |
| save_to_source_dir / custom_output_dir / output_filename | 输出位置与文件名 |
full_precision_mm=true 也能跑(仅磁盘变小)。weight_scale 反量化 FP8 源到真实值,避免出噪点。status (STRING):转换结果/错误信息。


binyuan Universal Quant Converter (NVFP4/FP8/GGUF) under category
binyuan/Advanced. Reads a source model from ComfyUI's checkpoints/unet/
diffusion_models, quantizes it to NVFP4 / FP8 / GGUF with a per-architecture
blacklist, and writes a new file (safetensors or .gguf). Preserves original
metadata and writes comfy_quant markers so ComfyUI loads quantized layers
correctly.pip install comfy-kitchen (NVFP4/FP8) and/or pip install gguf (GGUF).ComfyUI/custom_nodes/ComfyUI_binyuan_NVFP4_Converter/, restart.binyuan/Advanced → add the node.input_model.output_format: NVFP4 / FP8 (e4m3fn) / GGUF.model_type: Auto (Universal) (works on any architecture, recommended)
or a specific model (Flux/Krea2/Qwen-Image/Z-Image/Wan2.2/LTX/ERNIE/Ideogram/
LENS/PID/Boogu…).gguf_qtype.save_to_source_dir=true for same dir, or fill custom_output_dir
and output_filename.status.full_precision_mm=true (disk-only savings).weight_scale before re-quantizing (avoids noise).status (STRING): result / error message.__init__.py — registrationbinyuan_nvfp4_converter.py — node BinyuanNVFP4ConverterREADME.md / LICENSE (MIT)Binyuan NVFP4 Converter.git clone https://github.com/yjitiu/ComfyUI-Binyuan-NVFP4-Converter.git ComfyUI_binyuan_NVFP4_Converter