ComfyUI Runtime Precision Detector

Three simple ComfyUI nodes that show the precision ComfyUI is actually using at runtime.

• Detect Model Precision — MODEL to text
• Detect CLIP Precision — CLIP to text
• Detect VAE Precision — VAE to text

The report appears inside the node and is also available as a STRING output.

Active compute: BF16
Runtime weights: NVFP4
Quantization: NVFP4 (layout TensorCoreNVFP4Layout; logical BF16; storage UINT8)

Install

1. Download this repository.
2. Copy the ComfyUI-Precision-Detector folder into ComfyUI/custom_nodes.
3. Restart ComfyUI and refresh the browser.

No extra Python packages are required.

Use

1. Add a detector from utils/precision.
2. Connect a loaded MODEL, CLIP, or VAE.
3. Queue the workflow.

The detector reports active compute separately from weight storage. For example, an NVFP4 model that computes in BF16 shows both values.

Supported labels include FP64, FP32, FP16, BF16, PyTorch FP8 variants, FP4, NVFP4, MXFP8, INT8, INT4/ConvRot W4A4, mixed weights, and other PyTorch numeric dtypes exposed by the connected object.

An example is included in example_workflows/precision_detection.json. Select models installed in your own ComfyUI before running it.

Test

From the ComfyUI directory:

..\python_embeded\python.exe -m unittest discover -s custom_nodes\ComfyUI-Precision-Detector\tests -v

License

MIT