Detect CLIP Precision

Detect CLIP Precision is the text-encoder half of the pack's precision-detector trio. Feed it your CLIP object and it returns a plain-text precision_report showing the active compute dtype, the dtype of the loaded weights, any quantization, the manual cast policy, and the load/offload devices. No knobs, no side effects - it just reads the runtime state and tells you what your text encoder is actually doing. When you're running quantized or low-VRAM setups, "the encoder is in fp16" is an assumption you stop having to make.

How it works

The report is built by inspecting the CLIP object's internals at execution time: the compute dtype policy ComfyUI selected, the stored weight dtype (including whether the encoder is quantized - the labeler recognizes everything from plain FP16/BF16 up through FP8 E4M3FN, MXFP8, NF4, INT4, and ConvRot W4A4), any manual cast override, and the devices ComfyUI plans to run on and offload to. It's derived from the ComfyUI-Precision-Detector approach, reworked for the V3 extension API, and it prints the report on-screen as well as returning it as a string.

Input: clip, straight from your CLIP Loader / dual CLIP loader. Output: precision_report (STRING), and the node is marked as an output node so you see the text without wiring anything.

Why this matters right now: the text encoder is where modern models changed the most. SDXL's CLIP was a fixed, small thing; Flux and friends run a larger encoder that often gets quantized or offloaded to fit VRAM. The prompt-engineering story in the knowledge base is that the encoder swap changed everything about how prompts behave - so knowing what dtype and policy your encoder is running under is genuinely diagnostic when conditioning quality differs from expectations. If your encoder is quantized harder than you thought, that's the first thing to check before chasing prompt wording.

Installing

This is one of the 26 nodes in 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 it needs a current ComfyUI build; no extra Python dependencies, no model files.

Where people get burned

The report's honesty is also its limit: if the object doesn't expose the relevant attributes, the node prints UNKNOWN instead of guessing - so a line reading "UNKNOWN (not exposed by input object)" means "the loader didn't tell us," not "default." This is a runtime readout, so it reflects the policy ComfyUI picked for your current model and VRAM; it won't tell you what your checkpoint file was saved as. And it's an output node - branch the wire if the CLIP also needs to continue downstream, since the node returns only the report, not the CLIP.