Save Raw Latent (Creepybits)

Every time you tweak a VAE, change an upscaler, or want to try a different denoise strength, you normally re-run the entire sampler - minutes or hours of wasted GPU time just to get back to the same latent you already had. This node ends that. It saves the raw latent tensor straight from your KSampler's LATENT output, so you can decode the same generation a dozen different ways without ever resampling. Think of it as the "digital DNA" of your image, preserved before it's ever turned into pixels.

It's the saving half of a pair in the Creepybits suite: the companion Load Latent From Path node reads the file back and hands you the exact same LATENT to run through a VAE Decode. Save once, experiment forever.

How it works

The mechanism is brutally simple - and that's a compliment. The node grabs samples["samples"] (the raw tensor out of the LATENT dict) and calls torch.save() on it. That's it. No quantization, no metadata, no image encoding. What comes out is a .latent file containing the floating-point tensor exactly as the sampler left it.

It saves into your ComfyUI output directory, in a subfolder you name, with a timestamp appended so you never clobber the last save. Path looks like output/latents/MyLatentBlend_20260816-152033.latent.

The inputs that matter

• samples - the LATENT output from your KSampler. This is the whole point.
• folder_path - subfolder under the ComfyUI output dir, default latents.
• filename - the base name, default MyLatentBlend. The timestamp is added automatically.

There are no outputs - it's an output node, so it just does the save and reports the path in the console.

Installing it

Part of ComfyUI-Creepy_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/Creepybits/ComfyUI-Creepy_nodes

restart, and it appears under Creepybits/latent. ComfyUI Manager: search "Creepy_nodes". No models to fetch.

Where it bites

The big caveat is that you're saving a bare tensor with zero context. The file doesn't remember which checkpoint made it, what sampler ran, or what the prompt was - if you load a Flux latent into an SDXL decode, you'll get garbage and there's no built-in warning. Keep your own naming discipline (the filename field is your metadata).

The .latent format is a torch checkpoint, not an image - you can't preview it, and other tools won't read it. It's only useful with the pack's Load Latent From Path node (or a torch.load yourself). Worth it anyway: if you're the type who A/B-tests VAEs or upscalers, this single node removes the most expensive step from every iteration.