Detect VAE Precision

Detect VAE Precision answers the question that bites everyone running quantized or low-VRAM setups: is my VAE actually decoding in the precision I think it is? Feed it your VAE object and it returns a plain-text precision_report telling you the active compute dtype, the dtype of the loaded weights, any quantization, the output tensor type, and the load/offload devices. It's a diagnostics node, so it doesn't change anything - it just reads the runtime state and tells you the truth. If you've ever stared at banding in a decoded image and wondered whether the VAE was running in fp16 when you assumed fp32, this is the node that settles it.

How it works

This is one of three precision detectors in the pack (the others cover the diffusion model and the CLIP/text encoder), all derived from the ComfyUI-Precision-Detector approach and reworked for the V3 extension API. The VAE version reads a handful of real runtime attributes off the object: the selected precision policy (vae_dtype), the stored weight dtype, the quantization format if any, the output tensor dtype, and which devices ComfyUI plans to load/offload to. The report arrives as a string that the node also renders on-screen, because it's marked as an output node - you don't have to wire anything to see it.

The single required input is vae, straight from your VAE Loader, and the single output is precision_report (STRING).

What you're actually learning: whether ComfyUI is running the VAE with fp32 weights, a low-bit quantized layout (the report knows how to label things like FP8 E4M3FN, MXFP8, NF4, and ConvRot W4A4), or a forced cast policy. That's genuinely useful when you're comparing a fp16 VAE against a quantized one for quality, or when you're debugging why decode is slower than it should be (fp32 decode on a weak iGPU is a classic).

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 you need a current ComfyUI build; no extra Python dependencies and no model files.

Where people get burned

The honest limitation: the report is only as good as what the object exposes. If your loader hands over a VAE that doesn't expose a vae_dtype or runtime parameters, the node prints UNKNOWN for that line rather than inventing an answer - which is the correct behavior, but it means a "not exposed" line isn't the same as "running at default." Don't read this node's output as a performance benchmark either; it reports precision policy and device placement, not throughput. And because it's an output node, don't expect it to pass the VAE through - if your graph needs the VAE downstream, branch the wire, don't route it through the detector.