Load Latent

Load Latent reads a .latent file back into a LATENT so you can resume a workflow mid-stream. It's the inverse of Save Latent: instead of decoding a latent to pixels (and losing whatever the latent still holds that pixels don't), it loads the raw compressed tensor straight back into the graph. Think of it as a save-state file for your generation, not an image.

The one input is a dropdown, latent, listing every .latent file in your ComfyUI input folder. Files saved by Save Latent land in output/latents/ by default - so to reload one, move or re-point it into input/ first, and it'll appear in the list. Output is a LATENT you feed anywhere a latent would go: straight into a sampler for a continued pass, into a composite, into a decode.

Why you'd bother

Three honest uses. Resuming an expensive run: you spent twenty minutes on a long video latent, saved it, and want to try a different sampler or denoise on the same starting point without regenerating. Handing off latents between workflows without a lossy encode→decode→encode round trip - save once, reload in a different graph, keep the exact tensor. Debugging and sharing: a .latent is a compact, self-contained artifact (safetensors under the hood) you can share with someone so they see precisely the latent state you saw, not an image that's been through a VAE and a screenshot.

Mechanically it's simple but has one compatibility wrinkle worth knowing: the loader applies a scale factor based on a version marker in the file. Files written without the version flag get multiplied by 1 / 0.18215 - the classic SD 1.5 latent scale - so old-format files load with the scaling their model expects baked in.

The gotchas

The big one: a .latent is only meaningful to the model it was made for. It's a bare tensor with no model identity attached, and latent spaces are architecture-specific - a latent saved from an SDXL workflow fed into a Flux sampler produces nonsense, not a gentle degradation. If you're loading someone else's file, you need the same model family it was generated with. Second, remember it's not an image: the file stores the compressed latent and the channel layout, and nothing else about your settings travels with it - if you expect a full workflow restore, you want Save/Load of the whole prompt JSON, not this node. Third, the input-folder rule trips people constantly; the file simply won't be in the dropdown until it's in input/.

It ships with ComfyUI core - no install, no model files to download, just your own saved latents. For most people it's a rare guest in the graph, but when you need an exact latent state preserved, it's the only node that does it without a pixel detour.