Tensor Dtype Info

Tensor Dtype Info is the diagnostic node that ships with the SDXL 16ch Loader pack, and it's honestly useful even if you never touch a 16-channel checkpoint. You plug a MODEL, CLIP, or VAE into it and it tells you - as a text string - what precision that thing is actually running in. That's it. One input, one output, nothing else.

Why that matters: plenty of nodes quietly cast models to a dtype you didn't ask for, and "auto" precision is a black box. When your SDXL 16ch Loader (Precision) says bf16 but the image still looks off, or a pack loaded your model as fp16 and you want proof, this node settles the argument without digging through the Python console.

The inputs and outputs

Everything is optional, and you only feed it one thing at a time:

• model - a MODEL (the diffusion model / UNet).
• clip - a CLIP (text encoder).
• vae - a VAE (encoder/decoder).

The single output is STRING, one of fp32, fp16, bf16, plus a few extras like int8 for the rarer cases. Feed that string into a text-display widget or just read it on the node itself - it's a read-only utility, so it never blocks your render. Wire it wherever and leave it dangling; nothing downstream needs it.

How it works

Nothing clever under the hood. It grabs the first parameter of whichever module you passed in - model.model for a MODEL, cond_stage_model for CLIP, first_stage_model for a VAE - and reads its .dtype. One caveat worth knowing: it inspects the first parameter, not the whole model. For a uniformly-loaded checkpoint that's a fair answer. For a model that's been partially cast or fused into mixed dtypes, it's a peek, not a full audit. Don't ask it for a breakdown it can't give.

If you connect nothing at all, it returns the string "no input provided" instead of erroring, which is a nice touch for a debug node.

Install and gotchas

Same pack, same install as its siblings - ComfyUI Manager (search "SDXL 16ch Loader") or:

cd ComfyUI/custom_nodes
git clone https://github.com/NeuroSenko/ComfyUI-SDXL-16ch-loader

then restart. Zero dependencies. It lives in the utils category.

The one thing people trip on: the output is a plain STRING, not a display-friendly text box. If you want to see the value without adding a text viewer, just look at the little output widget on the node after it runs - the string shows there. And remember it's a diagnostic, so it works on any model, clip, or VAE in your graph, not just ones loaded through this pack. If your "precise" loader is silently ignoring you, this is the node that catches it lying.