MiniMax H3 Aicg 分段步骤

AICG3D_H3SegmentStep - MiniMax H3 Aicg 分段步骤 - is the other way to sample a context plan. Instead of one 分段采样 node swallowing the whole shot list, you place one Step per shot and chain them: step 1 → step 2 → step 3. Segment order is inferred from the chain, not from a widget.

Why would you want that? Control and cache. Each step is its own node in the graph, so ComfyUI caches each one, and you can override a single segment's prompt without rebuilding the plan. It's the fiddly-but-precise option; the Render node is the fast one. Both end up at the same place.

What happens in a step

The first step receives the setup object from 采样设置 and figures out it's segment 1. Every later step receives only the previous step, inherits its setup, and increments the index - the node literally computes index = previous.index + 1 and then checks it against the plan. From there it rebuilds that shot's conditioning from its own prompt and media, applies the Motion Context tail (or RGB guide, depending on the context's continuity mode) from the previous segment, samples, and passes itself plus its sample downstream.

The finished latent stays on CPU between steps - that's deliberate, it's how a long chain doesn't wedge the card - and the previous segment's tail latent comes along for the ride instead of being re-decoded.

Inputs and outputs

Required: seed. That's it, and it's the segment's own seed, so the control-after-generate widget applies per step.

Optional, and they're the whole interface:

• sample_setup - on the first step only. Connect 采样设置 here.
• previous_segment - on every step after the first. This is the chain.
• prompt_override - a string input that replaces just this segment's prompt. Handy with a text primitive: iterate the wording of shot 4 while everything else stays cached. If the override includes explicit <Picture N> / <Video N> / <Audio N> tags it can also reselect which media this segment sees; a plain text override keeps the media the context assigned, which is the behaviour you want most of the time.

Output: segment (MINIMAX_H3_SEGMENT_STEP). Into the next step, or - at the end of the chain - into MiniMax H3 Aicg 分段汇总, which collapses the chain into the segments object that 分段解码 and 分段二采 expect.

Install

Nothing special:

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

Restart, Ctrl+F5, look under AICG3D. You'll need a MiniMax H3 bundle loaded upstream and a context node that produced a shot list - this node doesn't invent segments, it consumes them.

The errors you'll actually see

"The chain contains more steps (5) than Context Segments (4)." You added a step the plan doesn't have a shot for. Delete it or extend the plan.

"Connect Segment Sample Setup only to the first Segment Step." Two setups in one chain. The second one is ignored by design; remove it before it confuses you later.

"The chained Segment Steps come from different H3 Context plans." You edited something upstream so the context rebuilt its plan, and now step 3 holds a reference to the old object while step 4 holds the new one. The identity check is strict on purpose - half-mixed chains produce subtly inconsistent output rather than an obvious error, so the pack fails loudly instead.

"The final step contains 3 segments; expected 5." That's 分段汇总 talking: it wants the last step, not whichever one you happened to grab. Wire the terminal step of the chain.

A structural note: because each step caches independently, re-running with only segment 6 changed re-executes segment 6 and everything downstream of it, and nothing upstream. That's the real argument for this node over Render, and it's a good one once you're in reroll hell.