JR MiniMax H3 Sequential Continuation Guide

Every chunked long-video scheme has the same problem: chunk 1 generates a person mid-gesture, and chunk 2 starts with a different person in a different room, because each chunk is a fresh generation. The usual fix is "just reuse the same prompt," and it's not enough - prompts don't remember poses. This node is the fix that actually works: it anchors each new chunk to the last committed frame of the previous one, using MiniMax H3's native image-guide mechanism. It's the second node in the pack's sequential audio chain, and it's what makes the whole thing look like one continuous take instead of four clips.

What it does

JR_H3_SequentialContinuationGuide sits between the Chunk Driver and your sampler. It takes the positive CONDITIONING, the latent, the chunk_context token from the Driver, and a VAE, and it has two modes:

• Previous Last Frame (the default) - chunk 1 uses the initial_frame you connect (required; it errors without one). Chunks 2+ load the previous chunk's committed terminal frame from the job's continuation folder on disk and apply it as a guide.
• Independent MV - passes the conditioning through unchanged. Every chunk is on its own visually; useful if you actually want cuts, or if a previous frame is missing.

The mechanism, grounded

The continuity trick is delegating to ComfyUI's native MiniMaxH3AddGuide node and applying it at local frame 0. The previous chunk's final RGB frame is placed as the first-frame image guide of the current chunk - so the sampler starts from where the last chunk actually ended, pixels and all. The guide is read from a PNG (continuation/last_00000.png etc.) that the Sequential Video Output node wrote when it committed the previous chunk, which is why this node works across separate ComfyUI prompts and even survives a browser restart: the state lives on disk, not in RAM.

Two honest limitations, straight from the pack docs. First, there's no hidden-state or KV carry between chunks - the guide is a visual anchor, not a numerically seamless handoff, so motion can still wobble at seams. Second, the guide must be applied before you build the Guider for the chunk - wire the node's output positive into your Guider, not the original conditioning.

JR_H3_SequentialAudioChunkDriver (chunk_context, latent)
  -> JR_H3_SequentialContinuationGuide
       positive (guided) -> Basic Guider -> Sampler
       latent ------------------------------> Sampler

Outputs: positive, latent, the same chunk_context, and a status string that reports which source the guide came from (connected initial_frame for chunk 1, the PNG path for later chunks).

Installing it

Same pack, same drill - JR MiniMax H3 (Goldlionren/ComfyUI_JR_MiniMaxH3Node), via ComfyUI Manager (search "JR MiniMax H3") or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Goldlionren/ComfyUI_JR_MiniMaxH3Node.git
<ComfyUI-Python> -m pip install -r .\ComfyUI_JR_MiniMaxH3Node\requirements.txt

Run that pip command with the same Python that runs ComfyUI - the classic miss with this pack. It needs imageio-ffmpeg>=0.5, and it expects a ComfyUI current enough to ship comfy_extras.nodes_minimax_h3 (day-zero H3 support). Restart after installing.

Troubleshooting

The errors here are honest and specific. "Previous Last Frame mode requires initial_frame for chunk 1" - you forgot the initial_frame connection. "Previous chunk continuation frame is missing: ..." - the previous chunk never committed its terminal frame (maybe the output node was bypassed), and the fix is either re-running the preceding chunk or switching to Independent MV. "MiniMaxH3AddGuide API is incompatible" - your ComfyUI is too old for the native H3 guide node. None of these are silent failures, which is exactly what you want in a node that decides visual continuity between minutes of expensive generation.