VAE Clock Start (Latent)-SG

If you've ever watched a video workflow grind away and wondered whether the VAE decode is the part eating your minutes, this is the front half of a stopwatch for exactly that question. VAEClockStartSG doesn't touch your image. It takes in a latent, notes the wall-clock time, prints a timestamp to your ComfyUI console, and hands the latent straight through. Nothing about your pipeline changes - you're just instrumenting it.

It only makes sense as half of a pair, though. Put this node before VAEDecode and its sibling VAE Clock Stop (Image)-SG right after it, and together they print how long the decode actually took. On their own they're a timestamp log in the terminal; together they're a poor man's profiler for the single slowest step in a video generation.

Why you'd reach for it: VAE decode is the bottleneck in video workflows. A single image decodes in a couple of seconds, but decoding a multi-second clip at high resolution can run for minutes - video models use a heavier 3D temporal VAE for a reason. The author built this pack precisely because decode is so much slower for video that it deserves its own per-workflow timer, rather than a global console clock stamped into every job. That's a fair take: image-only workflows don't need the clutter.

How it works

When the node executes, it records time.time() into a module-level dictionary keyed by the latent tensor's ID, and prints something like 🎞️[VAE Clock Start]🎞️ 2026-08-16 14:03:22 to the terminal where you launched ComfyUI. Then it returns the latent unchanged - the same object, byte for byte. Safe to add, safe to remove, zero effect on output quality, and it works with every latent type including tiled decode, since it's a pure pass-through.

One important detail: the console, not the canvas. Nothing this pack does shows up in the web UI. If you're staring at the graph expecting a badge, you're looking in the wrong place - the output lives in the ComfyUI terminal window.

Inputs and outputs

The whole thing is one input and one output:

latent (LATENT, required) - feed it the latent that's about to go into VAEDecode.
Output: LATENT - the same latent, untouched.

Wire it like this: Sampler output → VAE Clock Start → VAEDecode (latent input) → VAE Clock Stop → Save Image. There's nothing else to configure - no settings, no knobs, no time unit to pick.

Installing it

ComfyUI Manager is the easy route: open Manager → Custom Nodes → search for VAE Timestamp Clock (the pack title is "ComfyUI VAE Timestamp Clock SG") → install → restart ComfyUI and refresh the browser.

Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/ShammiG/ComfyUI-VAE-Timestamp-Clock-SG.git

Then restart ComfyUI. There are no dependencies to worry about - no requirements.txt, no model downloads, no torch extras. It's pure Python stdlib, so the install is as painless as custom nodes get.

Gotchas

The sibling Stop node grabs the most recent start entry rather than matching this latent by ID, so keep it to one decode at a time. And if your Stop prints "No start Clock found!", this node never executed - usually because nothing downstream actually consumed the latent, so ComfyUI's lazy execution skipped the whole branch. Trace your wires and make sure the image end of the chain lands in a Save Image or Preview node.