Nodes/ComfyUI-IcyHider/Icy SaveLatent
ComfyUI Node

Icy SaveLatent

Saving Latents for Later

By icekiub-ai·Created 9 months ago·Updated 9 months ago· 31
Icy SaveLatent
  • samples
  • samples
filename_prefixlatents/ComfyUI

IcySaveLatent writes your latent tensor to disk so you can load it back later with LoadLatent and pick up exactly where you stopped - no re-sampling, no decode-encode detour through pixels. It's the save-game feature of ComfyUI, and it's the node you reach for when a generation you like needs a second, third, or tenth experiment run off it.

The "Icy" prefix is the pack's privacy coat, not a different saver. IcyHider subclasses core ComfyUI nodes at startup - same inputs, outputs, and behavior, filed under the IcyHider Comfy Core category - and its JavaScript hides previews until hover. This is core SaveLatent with a cover on.

Inputs and output

From the schema, two inputs, one output:

  • samples - the LATENT to save, typically straight out of a KSampler or a VAE Encode.
  • filename_prefix - file name prefix, default latents/ComfyUI. The default folder structure means files land in ComfyUI/latent/latents/ - you can change the prefix to organize by experiment or date.
  • Output samples - the latent passes through, so you can save mid-chain and keep building on it.

When saving a latent beats saving an image

Two situations, both common. First, branching: you've generated a base you like and want to try five different second passes - upscale, different denoise, different prompt, different sampler. Save the latent once and each branch starts from byte-identical input. Re-rolling from a saved PNG works too, but every decode-encode cycle is a lossy trip through the VAE, and the KB's concepts essay is blunt that latent compression is lossy and the loss compounds. Latents skip that round trip entirely.

Second, continuity: you're building a workflow you'll reuse, and you want the "input" stage to be a stable saved state rather than a generation that has to be re-run each time. LoadLatent + SaveLatent gives you that.

The one rule that matters: a saved latent is married to the VAE and architecture that produced it. Save an SDXL latent and load it against an SD 1.5 VAE and you get noise, not a picture. Keep the pair with the same checkpoint family, and you're fine.

Install

Frontend-only pack, no dependencies:

# ComfyUI Manager: search "IcyHider", Install, restart

# or:
cd ComfyUI/custom_nodes
git clone https://github.com/icekiub-ai/ComfyUI-IcyHider.git
# restart ComfyUI

Pack-wide: the cover is DOM-based, so the newer Nodes 2.0 canvas frontend may not hide previews, and it's screen-only - the latent file on disk is real and anyone with folder access can copy it.

CategoryIcyHider Comfy Core

Inputs (2)

NameTypeDefaultDescription
samplesLATENT
filename_prefixSTRINGlatents/ComfyUI

Outputs (1)

NameTypeDescription
samplesLATENT