Detect Model Precision

Detect Model Precision is the flagship of the pack's three precision detectors, and it answers the question that haunts anyone running quantized checkpoints: is my model actually doing what I think it's doing? Feed it your MODEL object and it returns a report of the active compute dtype, the stored weight dtype, any quantization layout, the manual cast policy, and the load/offload device plan. It changes nothing - it's a pure readout - but it settles arguments about whether that GGUF or fp8 or ConvRot conversion is really taking effect.

How it works

At execution time the node inspects the model wrapper's runtime state: the inference dtype ComfyUI selected for compute, the dtype of the live weights, the quantization format if the checkpoint is quantized (the labeler is thorough - it names FP8 E4M3FN/E5M2, MXFP8, NF4, FP4, INT4/INT8, and ConvRot W4A4, along with storage vs logical dtype and kernel dtype when exposed), any manual cast override, and the devices ComfyUI will load and offload to. The logic is adapted from the ComfyUI-Precision-Detector project into the V3 extension API, and the result is both returned as a string and rendered on the node itself, because it's marked as an output node.

Input: model, from any checkpoint / UNET / model loader. Output: precision_report (STRING).

The genuinely useful reading: "Active compute" tells you what dtype the sampler math runs in (fp16 vs bf16 vs fp32), "Runtime weights" tells you what the weights are stored as, and "Quantization" tells you the actual quant layout. The gap between "compute" and "weights" is where a lot of real-world confusion lives - a model can store fp8 weights but compute in bf16, and the report is one of the few tools that shows you both at once.

Installing

Part of ComfyUI-Flow-Assistor - ComfyUI Manager (search "Flow Assistor") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Merserk/ComfyUI-Flow-Assistor.git

Restart after cloning. The pack is V3-only, so a current ComfyUI build is required; no extra Python dependencies, no model downloads.

Where people get burned

Remember this is a runtime policy readout, not a file inspection: it reflects what ComfyUI decided for your current model, VRAM, and settings, so the report can legitimately differ from what the checkpoint file declares. Lines reading UNKNOWN mean the object didn't expose that detail to the node - read them as "couldn't see," not "doesn't exist." And because the node is a terminal output, don't route your model through it expecting it to pass the model along; branch the wire if the model needs to keep flowing downstream.