ANDRO Range Stats

The question this node answers is the one every number in the comfyui-vae-float32 README was measured with: how much of a frame actually lives outside [0,1], and how finely is it quantised? Everything the pack claims about bf16 being coarser than you think comes from running this on real generations - a stock bf16 decode reading ~77 distinct values in the [0.2, 0.3] window against ~3.3 million for a float32 decode of the same latent.

It's a measuring node, and its one superpower is that it passes the image straight through untouched. The image output is byte-for-byte what went in, so you drop it in a chain, not beside it - decode → Range Stats → whatever's next - and read the numbers without adding a branch or disturbing the wiring.

What comes out

• report - a string with min/max/mean, percentiles, the share of samples outside [0,1] that the stock clamp would delete, and the quantisation step reported as effective bits (a bf16 decode reads as ~10.0, real 8-bit material as 8.0). When the distinct-value count nears the sample count in the window it says SATURATED, because past that point you're measuring the picture's sample count, not the format's precision.
• histogram - a log-scaled image with the [0,1] clamp limits marked in red. Log scale matters: the out-of-range tails are a fraction of a percent, so on a linear axis they're invisible - which is exactly why people believe nothing is lost.
• banding_mask - a MASK showing where banding would show up first. The reasoning is worth understanding: a band is one quantisation level held across several pixels, so you need a real gradient and local noise below the step. Noisy material dithers itself and never bands; clean sky and haze do. Wire this to a preview and you'll see precisely which parts of your shot need the precision.

Inputs are just image and a free-text label (so two of these in one graph can be told apart in the log). That's the whole node - two widgets.

The part that feels like a bug but isn't

Here's a ComfyUI trap that makes measuring nodes look broken: the strip that appears when you select nodes on the canvas has a ▶ button, and it does not mean Run. It runs Queue Selected Output Nodes, which tells the executor to keep only the selected output nodes and drop every other node - including Range Stats sitting mid-chain. The run still reports success, with no message about what got skipped. A plain Run (the big button) queues every output node and everything executes. If your stats node "never runs," check that you didn't hit the selection-play button.

Install

This is one of ten nodes in the comfyui-vae-float32 pack - install via ComfyUI-Manager by searching comfyui-vae-float32, or:

cd ComfyUI/custom_nodes
git clone https://github.com/AndreiOrehov/comfyui-vae-float32

Restart, find it under the ANDRO category. The starter workflow the pack ships, 01_measure_your_vae.json, is the fastest way to see the point: it round-trips one image through any VAE you own, decodes it twice (stock and float32), and runs two Range Stats nodes side by side. Same latent, same VAE, one readout with a few hundred thousand more levels than the other. No LTX, no video model, nothing to download - just point VAELoader at a VAE you already have and hit Run. It's by Andrei Orehov / Andromediastudio, Apache-2.0, and it costs nothing to try; whether float32 matters for your shots is exactly the question this node answers before you spend the VRAM.