MiniMax H3 Aicg 分段汇总

What it's doing in the chain

The AICG3D segment workflow renders long video shot by shot. Context Segments builds a multi-shot plan, then each shot gets a Segment Step node, and those steps are chained backwards: step 2 holds a reference to step 1, step 3 to step 2. That's how continuity survives the split - each segment can reach back through the chain for the previous shot's state instead of relying on a wire you have to keep plugging and unplugging as you add shots.

Which means the finished chain isn't a list sitting in one place. It's a linked structure whose tail is the only handle on the whole thing. MiniMax H3 Aicg 分段汇总 (AICG3D_H3SegmentCollect) takes that tail and walks it back into a plain collection that the decode node understands.

One input, one output, no settings. It is the least interesting node in the pack and you cannot skip it.

How it works

Input: final_segment (MINIMAX_H3_SEGMENT_STEP). Output: segments (MINIMAX_H3_SEGMENTS), which goes to MiniMax H3 Aicg 分段解码.

The collect routine follows previous pointers from the node you connected until it runs out, reversing as it goes so the result reads first shot to last. It does a proper job of refusing bad input:

Every link must be a Segment Step; a cycle is detected and rejected rather than hanging.
Each step's index must be sequential (1, 2, 3…) and all steps must share the same context plan. A chain stitched from two different plans fails.
The step you connected must be the final one. Connect shot 2 of 3 and it tells you it needs step 3 and that it got 2.

Then it merges the per-shot prompt and media overrides back into the plan's shot list, marks the stage as sampled, and emits the result.

The checks matter more than the merge. Every one of those errors is a graph mistake with a precise name, and they fire before you spend GPU time on a decode of something malformed.

Wiring

Context Segments ──> Segment Step 1 ──> Segment Step 2 ──> Segment Step 3
                                                                  │
                                                        AICG3D_H3SegmentCollect
                                                                  │ segments
                                                          Segment Decode

You connect exactly one wire, from the last step in the chain. Not all of them. Not the first. The node's own description is unusually blunt about this, which suggests how often people got it wrong.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/JGRFW/comfyui-AICG3D

Restart, Ctrl+F5. Manager install works too - Registry search AICG3D or 短剧工作台. requirements.txt is imageio-ffmpeg only; nothing here needs a model file of its own, though the chain it feeds obviously does.

You don't need to add this node to an existing graph by hand if you're using the pack's 13-series workflow - it's already wired. It's when you're building your own segment chain that you'll go looking for it.

Traps

Wrong step connected. The error names the count and the index, so read it rather than rewiring at random.

A broken chain in the middle. If you deleted a Segment Step node from the middle of the chain, the indices no longer line up and collection fails with a continuity complaint. Rebuilding forward from the last good step is faster than trying to repair the chain.

Two plans in one chain. Shot overrides are stored per-plan. Mixing contexts - say, half your shots from a re-run Context Segments node - produces a chain that can't be collected. The pack's own fix for this class of problem is to rebuild the whole run rather than splice.

Expecting it to decode. It doesn't. segments is a plan-with-samples; the decode step is a separate node, and that separation is what lets you collect first and decode once, at a resolution you choose after seeing the first pass.