Load Diffusion Model (Quantized)

This is the QuantOps replacement for Load Diffusion Model - the node you use when your diffusion transformer is a separate quantized file in diffusion_models, and you want to pick exactly how it's interpreted. The whole point is the quant_format dropdown: the file might be int8, tensorwise int8, one of three fp8 layouts, mxfp8, hybrid mxfp8, or nvfp4, and this node makes sure ComfyUI builds the right ops for whatever's actually inside.

The reason this matters more than it sounds: since Flux, "the diffusion model" is one file and "the text encoder" is another, and they're quantized independently. The KB's guidance is to spend your precision budget on the diffusion model and squeeze the encoder - so if you're running a quantized UNET at all, you've already decided the encoder takes the bigger hit. This node is the half of that equation that gives you control.

How it works

Mechanically it's the same core as the other full loaders: read the safetensors, detect (or force) each layer's quant format, package that as quantization metadata in ComfyUI's model options, and attach the pack's custom kernels. auto scans the file and matches layer formats; the explicit options are for when you know a file's layout and want to pin it. The kernel_backend choice routes int8 matmuls through either plain PyTorch or the triton kernels.

The inputs that matter

• unet_name - the quantized diffusion model from your diffusion_models folder.
• quant_format - auto, or one of int8, int8_tensorwise, float8_e4m3fn, float8_e4m3fn_blockwise, float8_e4m3fn_rowwise, mxfp8, hybrid_mxfp8, nvfp4. Leave it on auto unless you have a reason.
• kernel_backend - pytorch or triton. Triton is the fast int8 path; it also carries the biggest setup burden.
• disable_dynamic and low_memory - the pack's shared loading toggles.

Output is a single MODEL, straight into your KSampler. No CLIP, no VAE - those are separate loaders for separate files.

Installing it

ComfyUI Manager → search "ComfyUI-QuantOps", or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-QuantOps

Restart and let requirements.txt install unifiedefficientloader>=0.5.2. Quantized UNETs live on silveroxides' HuggingFace.

The caveats

The fast int8 path is a checklist, not a toggle: for int8_tensorwise on an RTX 30-series card you want torch 2.10+cu130 or newer, the matching triton-windows wheel, silveroxides' pre-compiled comfy-kitchen int8 wheel, and the --enable-triton-backend launch flag - otherwise the fast backend won't engage. And the pack is deprecated: the README opens by announcing that int8 ConvRot support is now built into ComfyUI, the author stopped maintaining, and doesn't want low-effort issue reports. On a current ComfyUI the native Load Diffusion Model handles the modern formats; this node's useful life is loading older quantized UNETs produced by the convert_to_quant toolchain.