Minimax_H3_AutoContext_parameter

If you grabbed the AutoContext sampler and wondered where the prompt box is - it's here. This pack splits the work: H3Parameter is a pure configuration node, a dict-builder that bundles your prompt, resolution, frame counts and audio settings into a single parameter output. That dict is the sampler's required input. One parameter node, one sampler, nothing else to wire.

It also shows a live "expected segmentation" preview on the node itself. That's a front-end JavaScript calculation, not inference - it just shows roughly how your video will split into chunks before you burn GPU time. Genuinely handy when you're tuning, and worth knowing it won't tell you anything about quality.

The inputs that matter

• long_prompt - the prompt. This is the one you'll spend all your time on.
• total_frames / chunk_frames / context_frames - the heart of the chunking. Defaults: 362 total (~15s at 24fps), 90 per chunk, 22 context. MiniMax H3's VAE works on a 17-frame temporal grid, so valid counts are 5, 22, 39, 56, 73, 90… The node snaps your numbers onto that grid, so don't fight it. Context frames are the previous segment's tail that carries into the next one - bigger means a smoother relay, and 22 is a good starting point.
• fps - only used for audio sync and converting the seconds in your prompt to frames. 24 by default.
• width × height - first-pass resolution (960×544 default). On a second pass the sampler ignores these and follows the input latent instead.
• lock_audio - on by default. During a second pass, only video gets re-sampled; the audio from pass one is reused. audio_drive flips it around: pair it with a drive_audio connection on the sampler and the video follows that specific soundtrack instead of generating its own.

The timeline stuff

prompt_mode has four personalities. auto (default) reads time markers like 0-5s in your prompt and slices accordingly, while untagged lines (style, soundscape, negative-ish notes) get spliced into every window. timeline forces the same, stricter. global dumps the entire prompt into every window - honestly the right choice for a one-shot homogeneous scene like a looped background or a talking head, and it sidesteps the trickiest limitations below. sequential lays lines out in order, with 全局:/[global] lines shared across all windows. If you're new to this, use global until you actually need per-scene direction.

clip_mode switches between Clip_Frame (even time slicing) and Clip_Tag, which splits by your own labels like 段1 / 段2 (the tag must end in a number, so A01 and [片段001] work too). Each tag becomes a full chunk; its duration comes from the tag's time range, then the segment's internal time markers, then a total_frames fallback. That's your shot-list mode.

Two limitations to keep in mind - the README calls them the node's constraints, and they only bite in segmented modes. First, temporal exclusivity: each chunk's prompt must describe only what's new relative to the previous chunk's end, because that end is the implicit starting state. Rewriting the previous chunk's action in the next one causes conflict and jittery motion. Second, references aren't remembered: if chunk 2 uses image1, chunk 2's prompt has to declare image1 again, or it runs without it.

Install and outputs

Same install as the whole pack: ComfyUI Manager, search ComfyUI_MinimaxH3_AutoContext, or git clone https://github.com/supElement/ComfyUI_MinimaxH3_AutoContext.git into custom_nodes and restart. No extra dependencies for this node itself - the pack's requirements only add scenedetect and opencv-python-headless for the seam-correction sibling.

The output is one parameter dict, and it's the sampler's required parameter input. That's the entire job: this node doesn't generate anything, it just makes the main node much less cluttered. Also note the pack downloads no model - H3 itself comes from ComfyUI's native MiniMax H3 support, and if you're in the US, EU, UK or Korea, the H3 Community License doesn't grant you rights to run the local weights there. Worth checking before you invest an afternoon.