Icy LoadLatent

IcyLoadLatent loads a latent you saved earlier with SaveLatent, bringing it back into the graph exactly as it was - no re-sampling, no decode-encode round trip. It's the "save game" button of ComfyUI workflows, and it exists because latents are the compressed representation the diffusion model actually works in. Saving one is like keeping your place in a book instead of re-reading the whole thing.

The "Icy" prefix is the pack's privacy layer, not a different loader. IcyHider subclasses core ComfyUI nodes at startup - same inputs, outputs, and behavior, moved into the IcyHider Comfy Core category - and its JavaScript covers previews until hover. So this is core LoadLatent with a coat on. One detail worth knowing: latents don't render a meaningful thumbnail anyway (they're noise-like tensors, not images), so the hiding here is mostly cosmetic - but if you're in Avalanche mode, this node plays along like everything else.

What it does

From the schema, one input, one output:

• latent - a dropdown of files in your ComfyUI/latent folder. That's where SaveLatent writes them, so this node is its mirror image.
• Output LATENT - the loaded tensor, ready to feed a KSampler, a VAEDecode, or another latent operation.

That's it. Pick a file, get your latent back.

Why save latents at all?

The honest uses: checkpointing mid-workflow and running branches. If you've got a base generation you like and want to try five different second-pass experiments off it, save the latent once and load it in five workflows - each one starts from identical input instead of re-rolling and hoping the seeds line up. It's also how you share "state" between separate workflow files without shipping images around.

The subtlety, and it's a real one: a latent is only valid for the VAE and architecture that produced it. Load an SDXL latent and try to decode it with an SD 1.5 VAE and you'll get garbage - different VAE, different latent space, as the KB's concepts essay drives home. SaveLatent and LoadLatent are a matching pair for a reason. Keep them paired with the same checkpoint family.

Install

Same frontend-only pack, no dependencies, no model downloads:

# 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 gotchas: the cover is DOM-based, so the newer Nodes 2.0 canvas frontend may not hide previews, and the hiding is screen-only - the latent file on disk is untouched. And if you change Avalanche or hide mode, ComfyUI needs a page reload for it to stick.