Prepared LTX · Decode Frozen Latent (T8 EXP)

Three nodes, three jobs: build the identity manifest, run the refinement, decode the result. This is the third one, and it's the only one in the trio that produces something you can look at.

The split exists because decoding is the part that most often kills a 16 GB card. Isolate it, and a decode crash doesn't cost you the generation - you reattach with the load node and try again. Isolate it, and you can also decode someone else's frozen run without running a single sampler.

What it does

It runs only the LTX VAE and movie worker, from a matching committed latent receipt. It outputs:

• video - an actual VIDEO output you can wire into a save or preview node like any other video. This node is flagged as an output node, so it shows up in the queue's output handling the way a save node does.
• saved_path - the file the worker wrote.
• report_json - the decode report.

The audio side is worth calling out because it's the thing people assume is broken when it isn't: the original audio is retained by the existing worker. In the pack's own qualification runs, the original AAC packets were kept and the decoded audio matched the input. Re-encoding to AAC is not bit-exact to PCM, so don't A/B those two and conclude something changed.

And the headline guarantee from the node's own description: generation is never implicit. It will not notice you didn't wire a generation receipt and quietly generate one for you.

Inputs that matter

• prepared_bundle - the manifest, same one the Generate or Load node used. Identity has to line up.
• generation_receipt - the typed output of either Prepared LTX · Generate Refined Latent or Prepared LTX · Load Frozen Refined Latent. Either works; that's the whole design. Generate means "I just ran it," Load means "I'm resuming."
• serial_lease_path (advanced) - absolute shared lease path for the worker's CPU/GPU leases, so two Core processes don't both grab the same resources.

Wire a Bundle → (Generate or Load) → Decode and you have a split pipeline. Wire Bundle → Load → Decode, with the load pointed at a SHA you generated yesterday, and you have a decode-only graph that touches no sampler at all.

Install

Manager, search MiniMax H3 Audio T8, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Restart ComfyUI completely after install or update - the pack's docs are insistent about this, because Python changes don't reload live. Then refresh the browser page.

The node pack itself installs nothing; the requirements.txt is comments explaining why. The heavy lifting is external and pinned: you provide the LTX-2 source, the Tao worker, the adapter weights, and a Prepared Python environment with nvidia-ml-py (pynvml) available. The environment the author qualified was Python 3.12 on a single RTX 4060 Ti 16 GB with 128 GB of RAM - worth reading as "what was tested," not "what will work."

Where people get burned

Handing it a receipt that doesn't match the bundle. Mismatched identity is an error, not a warning, and it's supposed to be.

Leaving the shipped example values in place. examples/workflows/61-prepared-ltx-split/ ships Generate_Decode and DecodeOnly graphs, and both say the bundle path and lease path are placeholders you must replace. The decode-only graph additionally needs the exact SHA-256 from the generation receipt you're resuming - the graph can't guess it, and it shouldn't.

Assuming long-video behaviour that hasn't been qualified. The pack's own evidence for this route is a 73-frame, 832×480, 24 fps indoor two-person clip that decoded completely with its original audio intact. That's one mechanical input case, explicitly not "arbitrary materials, providers, long durations or human quality acceptance." If your run differs, you're outside what anyone has proven - which is normal for an EXP node, and worth knowing before you spend an hour chasing a "bug" that's just untested territory.

If decode dies mid-VAE with nothing useful in the log, that's the failure mode the split was designed for: fix the environment, wire Load → Decode, and the generation you already paid for is still on disk under its hash.