IO_loadLatent

ComfyUI will happily save an image or a video, but a latent? That's the rare trick. IO_loadLatent is the reader for this pack's latent checkpoint system: it loads an H3 video/audio latent back off disk as a proper LATENT tensor, so you can pick up a generation exactly where you left off - same motion context, same audio latent - without re-running the model. It's the "load game" half of a save/load pair with IO_SaveLatent.

Why would you want that? The pack's whole multi-shot approach leans on latent continuity - chaining a previous clip's latent into the next generation so the character and motion persist. IO_loadLatent is how you make that persistence survive a ComfyUI restart or move between workflows: save the latent once with IO_SaveLatent, then load it anywhere later instead of regenerating or trying to reproduce it from an image.

Inputs

• latent_path (STRING, default "h3_context") - the file or folder to load. Relative paths are resolved from the ComfyUI output folder (that's where IO_SaveLatent writes). Point it at a file to load that file, or at a folder and use clip_index to pick within it.
• clip_index (INT, default 0) - which clip slot to load. 0 loads the newest .safetensors file in the folder, which is the convenient default; set it to a specific slot number to pin a particular clip.

Output: Latent (LATENT) - the loaded H3 AV latent, ready to feed a sampler or an AD_Inject_Latent-style pipeline.

How it works

It resolves the path, reads the .safetensors file, and reconstructs the joint AV latent as a NestedTensor with the video and audio members in the order the H3 stack expects. It validates that the file actually contains both - if it's not an H3 video/audio latent, you get an explicit "not an H3 video/audio latent" error rather than a confusing shape mismatch downstream.

It also watches the file: IS_CHANGED keys off the file's modification time, so if the file on disk changes, the node re-executes and picks up the new content instead of serving you a cached old latent. That's the behavior you want from a loader but rarely get.

When you'd use it

• Resume: save a mid-piece latent, come back tomorrow, load it and continue without continuity drift.
• Pass-off between workflows: workflow A generates and saves the context latent; workflow B loads it. This is how you keep a long project from becoming one giant graph.
• Debugging: when a multi-stage run gives you a weird jump, load the saved latent and inspect what the previous stage actually produced.

One honest caveat: it only reads the format IO_SaveLatent writes (metadata tag h3_motion_context_av_v1), so it won't load arbitrary latents from other packs. Use the pair together and it's seamless.

Installing it

From cardenluo/ComfyUI-Apt_Preset:

cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset
pip install -r requirements.txt   # or install.bat on Windows

Restart ComfyUI; it's under Apt_Preset/IO_Port. No model downloads needed.