MiniMax H3 Aicg 采样设置

AICG3D_H3SegmentSampleSetup - MiniMax H3 Aicg 采样设置 - is the least glamorous node in the pack and the one that keeps the graph from becoming a spider. If you're doing the per-segment 分段步骤 style workflow instead of the all-in-one 分段采样, you'll need it. Otherwise you can ignore it entirely.

The problem it solves: every segment step needs the same context, model, sampler and sigmas. Wire those into each of twenty steps and you have eighty wires, any one of which can be wrong in a way that produces plausible-looking output you'll spend an hour debugging. So the pack lets you connect them once, on this node, and thread a single setup object through the chain instead.

What it does, mechanically

It bundles five things - the H3 context, the resolved model, the sampler, the sigmas, and an optional sampling strategy - into an immutable object, and hands it downstream. That's it. There's no sampling here, no latent, no image. The step node that receives it carries the same object forward through the chain, which is why later steps need nothing but a previous_segment.

Two validations fire at build time, and they're worth knowing before you hit them:

• If the connected context is a selected-video plan (built by 已选视频上下文), it refuses: "Segment Sample Setup supports generated Context Segments, not selected-video preparation." The selected-video path takes its own route into Segment Sample.
• If you attach a SelfLift-style strategy while the context's continuity mode is one of the AV modes, it refuses too. SelfLift supports latent and RGB-guide continuity only, not the soft/hard AV variants.

Inputs and outputs

Required: h3_context, sampler, sigmas. Optional: model and sampling_plan.

• h3_context must come from a context node that actually produced a shot list. An empty context errors with "The connected H3 Context contains no segment plan".
• model is optional because the context usually carries the H3 bundle and its model. You only pass one explicitly when you're routing a differently-patched model - an accelerated or distilled one, say - down the chain.
• sampling_plan is the socket for AICG3D_H3SelfLiftStrategy. Optional, and if you don't connect anything, sampling is a plain denoise with your sampler and sigmas.

Output: setup (MINIMAX_H3_SEGMENT_SAMPLE_SETUP), which goes into the sample_setup input of the first 分段步骤 node. Only the first. The step node throws a specific error if you also connect it to a second one, because the setup is supposed to travel with the chain.

Install

Same pack, nothing extra:

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

Restart and hard-refresh. It lives under the AICG3D category; the node label is Chinese, so search AICG3D or Aicg in the node search rather than the class name.

When it goes wrong

The failures here are all wiring, and the error messages are unusually blunt, which I appreciate:

• "Connect Segment Sample Setup to the first Segment Step." You started the chain with a step that has no setup and no previous segment. Something has to be first.
• "Segment Sample Setup supports generated Context Segments, not selected-video preparation." Wrong context node feeding this one - swap to the regular context segments node, or use the selected-video pipeline's own entry.
• "The connected H3 Context contains no segment plan." The context node hasn't been run yet, or its own inputs failed. Run it or check what it's complaining about upstream; this node can't build a plan of its own.

One thing this node will not do is change settings per segment. Everything in the setup object applies to the whole chain. If segment 12 needs different sigmas from segment 11, you're in the wrong workflow shape - use 视频段落 nodes, where each segment owns its own prompt, seconds and seed.