Load MiniMax H3 AV Latent

So you saved a MiniMax H3 latent with Save MiniMax H3 AV Latent and you want it back. The temptation is to grab the stock Load Latent node and point it at your .h3latent file - and that's where it falls apart, because H3 latents aren't plain tensors. They're NestedTensor objects that bundle a video latent and an audio latent together, and most of ComfyUI's built-in latent nodes don't speak that language. Load MiniMax H3 AV Latent is the counterpart that reassembles them correctly.

How it works

This is the load side of the MiniMax H3 save/load pair. It reads a .h3latent file from ComfyUI's input folder - these are safetensors files under the hood, with each latent member stored separately as latent_0 (video) and latent_1 (audio), plus a tensor_count header so the loader knows how many members to expect. It validates that header, then rebuilds the latent: a single-member file comes back as a plain tensor, and a file with multiple members is reconstructed as a comfy.nested_tensor.NestedTensor. What you get out is a proper LATENT that the rest of your graph can consume.

One nice detail: the node hashes the file to decide whether it needs to reload, so re-running a workflow doesn't waste time re-reading an unchanged latent.

The one input

• latent - this isn't a value you type; it's a dropdown listing every .h3latent file in your input folder. Pick your file, done.

That's it. No filename strings, no subfolder browsing, no other knobs. The single output is LATENT, which you wire into whatever comes next - a sampler for further denoising, or straight into a VAE decode to render frames and audio. If you saved a latent mid-pipeline and want to keep iterating from that exact point, drop this node at the front of your workflow in place of the model-driven encode and you're back where you left off.

Installing it

Same drill as the save node - it ships in the same pack, and the pack has zero external dependencies. No models to download, no API keys, nothing to configure. Either install via ComfyUI Manager (search "MiniMax H3 Latent") or:

cd ComfyUI/custom_nodes
git clone https://github.com/JerryZRic/comfyui-minimax-h3-latent

Restart ComfyUI and look for both nodes under model/latent/minimax.

Gotchas worth knowing

The big one is folder placement. The file picker only lists .h3latent files sitting directly in the input folder - no recursion, so anything in a subfolder won't appear in the dropdown. Drag your files to the top level and restart the node's list if they don't show up. Also, this loader only understands files written by the pack's own save node; a latent saved by the stock Save Latent won't have the tensor_count header and will be rejected with a clear error. And since this pack is days old (first commit mid-2026), expect a little rough-around-the-edges energy - if something's off, the GitHub repo is small enough to skim. Use the save node, keep files at the input root, and this pair quietly does the job the built-in latent nodes can't.