SP_LoadLatent

A latent is the compressed, noisy-looking tensor a diffusion model actually works in - not the finished image, but what a VAE decodes into one. Normally a latent only exists transiently, mid-workflow, between your sampler and your VAE Decode. SP_LoadLatent is for the case where you've saved one to disk and want it back exactly as it was - reproducing a specific intermediate state, resuming a workflow from a checkpointed step, or feeding an exact latent (rather than a freshly-generated one) into a downstream node without re-running everything upstream.

How it works

It's about as simple as a loader gets: one field, filename, which takes an absolute path to a latent file on disk - the default placeholder value literally says abs_path_to_latent..., which is the node telling you not to bother with a relative path or a bare filename. Point it at the file, and it outputs a LATENT you can wire anywhere a normal sampler's latent output would go - into another KSampler for further denoising, straight into a VAE Decode, or into any other node in this pack (or ComfyUI itself) that expects one.

The node doesn't save latents itself - it's the read half of a save/load pair. You'll need something upstream, elsewhere in your workflow or from a previous run, that wrote the file in the first place; ComfyUI's own core nodes include a latent-saving node for exactly that purpose.

Inputs and outputs that matter

• filename (STRING) - the absolute path to the .latent file you want to load. No relative-path support implied by the default placeholder, so use the full path.
• Output: LATENT - the loaded tensor, ready to wire into any sampler or VAE Decode.

How to install it

Through ComfyUI Manager: search "SP-Nodes," install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/bananasss00/ComfyUI-SP-Nodes

Restart ComfyUI once it's cloned. No models, no dependencies - this is a file I/O node, not something that touches your checkpoint.

Common issues & troubleshooting

The node errors saying the file can't be found. Almost always a relative path where an absolute one was expected - double-check you're pointing at the full path (starting from your drive root or /), not just a filename or a path relative to ComfyUI's working directory.

Loaded latent doesn't decode to the image you expected. A latent is tied to the VAE architecture that produced it - a latent saved from an SDXL pipeline won't decode sensibly through a Flux VAE, or vice versa. If the decoded image looks like noise or garbage, check that the VAE you're decoding with matches whatever generated the latent in the first place.

Works fine locally but breaks after moving the workflow to another machine. That's the cost of an absolute path - it's tied to wherever the file physically sits, so a workflow shared between machines (or between a local machine and a remote/cloud instance) needs the latent file present at that same exact path on whichever machine is running it, or the filename updated to match.