📋 Latent Metadata Viewer v2.4

Before you commit a 20-minute decode to a latent file you're not sure about, spend 10 seconds here. Latent Metadata Viewer is the pack's read-only inspection tool: pick a file from the output/latents/ dropdown and it tells you everything about it without you loading a tensor into your graph.

The point is that a video latent is a big, opaque blob. Files easily run into the gigabytes, and while the tensor shape is technically printed in the console when you save, nobody remembers that an hour later. This node is the sanity check: which seed, which model, what shape, what format - all before you wire anything into a decoder.

It does this by reading the file's metadata rather than decoding it. Safetensors files store key-value metadata alongside the tensors, so the viewer opens the file, pulls the tensor shapes, and reports whatever the save node recorded - seed, model, timestamp - plus the format it actually found (safetensors or pickle). It never runs a VAE decode, so it's effectively free, and because it's an output node it stops the queue and prints its report to the UI.

All five outputs are strings: full_report is the whole thing formatted for humans, then shape, seed, model, and format are broken out separately in case you want to wire them into a text node or some condition logic. Inputs are just latent_file from the dropdown and refresh_list to re-scan. One gap worth knowing: unlike the pack's Load Latent (Advanced) node, there's no manual path input here - it only sees files inside output/latents/.

Install it the same way as the rest of the pack - ComfyUI Manager → search "ComfyUI-Grok-SmartVAE", or:

cd ComfyUI/custom_nodes
git clone https://github.com/uczensokratesa/ComfyUI-Grok-SmartVAE.git
cd ComfyUI-Grok-SmartVAE
pip install -r requirements.txt

Real requirements are torch, numpy, imageio, imageio-ffmpeg, psutil and opencv-python; no model downloads, no keys. It's the lightest node in the pack, and it shows - the whole thing is a dropdown and a report.

Practical gotchas are few. If seed or model come back empty, the file was likely saved without metadata, or loaded through the pickle fallback where nothing is guaranteed. If the dropdown is missing files you know exist, hit refresh_list. That's about it - this is a utility, not a workflow centerpiece. But the next time you're staring at ten .latent files trying to remember which seed was the good one, you'll be glad it's there.