MiniMax H3 Previous Context / 读取上一段上下文 (EXP/T8)

If you've ever chained H3 segments and watched the second one start a brand-new scene, you know the problem: H3 has a fixed generation window, and nothing carried over between runs. MiniMaxH3LongVideoContextLoadT8 is the fix. It loads the validated AV tail saved by the previous segment and hands it to the conditioning chain as head context, so segment 2 starts from where segment 1 ended instead of inventing its own world.

The clever bit is in how deterministic it is. Segment 0 returns an empty passthrough context - there's no previous segment, and the node doesn't pretend otherwise. And for segment N, it loads only the validated tail saved for segment N-1. It never picks "the newest file in the folder," which means rerunning a segment gives you a stable, reproducible context instead of silently gluing on whatever happened to be saved last. For a chain where you're iterating on segment 3, that determinism is the whole point - the head context won't change under you.

The inputs you set

Two inputs, that's it:

• chain_id - the string that names your chain. It must match the ContextSave and Planner nodes, or the load node will look for the tail in the wrong place and find nothing.
• segment_index - which segment you're about to render. 0 gives the empty passthrough; anything above it loads segment index - 1's tail.

That's genuinely all. The node is deliberately boring - it's the memory cell, not the brain.

What comes out

Three outputs. context (type H3_T8_CONTEXT) is the AV tail for the previous segment, wired into the Long Video conditioning node so H3 sees it as context. has_context is a boolean you can use to gate logic - segment 0 says false, later segments say true. report_json gives you the author's validation report so you can see what was loaded and whether the fingerprint matched.

Install and where it fits

Same shared story as the rest of the pack: install once via ComfyUI Manager (search MiniMax H3 Audio T8), or clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into ComfyUI/custom_nodes/ and restart. No pip extras, no auto-downloaded weights - you bring the H3 model files. It's marked Experimental and sits under T8/MiniMax H3/Long Video/Experimental, and the pack is GPL-3.0.

Use the pack's own workflow in 04-long-video as your starting point rather than hand-wiring this: the planner computes the segment window, this node loads the prior tail, ContextSave writes the new tail, and they all talk through chain_id. People get tripped up when the save node isn't run (or is run with save_context off) for segment N-1 - then N loads nothing, and you get a clean restart that looks like the feature is broken. It isn't; the previous segment just never saved its tail. Check report_json on both ends and you'll spot it in a second.