H3 T2VA Continue

This is where the review gate pays off. You looked at the preview from H3 T2VA Draft, you're happy with it, you press GO - and this node runs the remaining sampling steps of that same generation, decodes the joint video and audio latents and hands you a finished VIDEO you can wire straight into Core's SaveVideo. One node from approval to file.

It's the integrated counterpart to H3 Continue Sampler, and the difference is where the decode happens. The generic sampler pair stops at LATENTs and leaves decoding to your graph. This one owns the whole tail: video decode, audio decode, CreateVideo at 24 fps with the audio muxed in.

Inputs and outputs

Three inputs, two of them automatic:

• draft_state - from H3 T2VA Draft. It has to be that exact node's output on that exact wire; if you detach and reattach it, the old approval is void and you need a new Preview.
• go - the boolean behind the GO control. Disabled in the UI until the connected Draft has a reviewed Preview.
• approved_state_id - "Filled by GO. For API use, copy state.state_id from the Draft report." Only touch it when you're queueing from a script.

Outputs: video (a VIDEO - send it to SaveVideo, or any node that takes one), final_av_latent (the completed joint latent, if you want to decode or re-encode it yourself, and useful when you're debugging what the model actually produced), and report.

What "continue" actually means here

No new noise, no re-encode, no model swap. The node re-verifies the draft first - session, approval ID, upstream graph hash, model/LoRA runtime signature, payload fingerprint - then samples the remaining slice of the original schedule, with the noise object disabled and the original seed preserved. The report says it in fields rather than prose: resumed_from_step: 3, new_noise: false, image_reencoding: false, t2va_preserved: true.

Because the decode and encode happen inside this node, it is an output node - it will run and produce a video even if you haven't wired a Save yet. Handy for a quick check, mildly surprising the first time.

There's also a useful integration detail: when you queue with GO off, Continue returns an execution blocker instead of raising, so the whole downstream Decode → Save chain is skipped rather than erroring. That's what makes "queue with GO false, review, queue again with GO true" a safe habit instead of a minefield.

Install

Nothing extra to install - no pip packages, no model files, no downloads; the pack is pure node code. Manager search "MiniMax H3 Draft Continue", or:

cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/MiniMax-H3-T2VA-Draft-Continue
# restart ComfyUI

You need Core with native MiniMax H3 support and a recent frontend for the review panel. The author's validated environment is ComfyUI 0.37.0 with Frontend 1.52.7 on an RTX 5060 Ti 16 GB; the endurance run behind the current release did five full Preview → GO → Decode/Save cycles with no cumulative VRAM growth worth mentioning.

Where people get burned

Queuing twice doesn't add steps. A draft resumes the remaining schedule and then it's spent. If you want different motion, generate a new Preview - changing the seed or prompt after approval is exactly the case the stale check exists to catch, and it will refuse.

"This draft has not been reviewed. Preview again, then press GO for that exact draft. The cached state may have expired." The usual cause is one of three: you restarted ComfyUI, you reloaded the browser tab, or you opened a saved workflow. Approval state is deliberately session- and page-local, so Ctrl+F5 throws it away. Nothing is broken; Preview again.

A Continue node with no state attached. Feed it the output of the generic H3 Draft Sampler and the pack tells you which state type it wanted. The two pairs look interchangeable in the node search and are not.

The whole run fails after the video already looked done. The pack only reports COMPLETE once the downstream output chain succeeds, so a failing Save node shows up as a review error rather than a green tick you can't account for. If your queue is empty and the status panel says otherwise, look at the downstream node, not at the sampler.

Changed the model or a LoRA after Preview? The runtime signature won't match and GO is refused with a message telling you to generate a new Preview. It never quietly substitutes a patched model - that's the most valuable thing about this whole pack, honestly, and the reason to prefer it over rolling your own split sampler with a couple of KSamplers in a row.