MiniMax H3 Aicg 无限段落顺序生成（全局设置）

AICG3D_H3SequenceGlobal - MiniMax H3 Aicg 无限段落顺序生成（全局设置） - is the node you wire first and touch least. One config, fed to the first 视频段落 and inherited by every segment after it: canvas, handoff, sampling, VRAM behaviour, cleanup. It's the film's spec sheet, not a per-shot control. You set 480P or 720P once and twenty segment nodes stop arguing about it.

The settings that actually change your output

resolution (360P through 1080P) and aspect_ratio (1:1 to 21:9) together compute the canvas; the resulting width/height is shown read-only under them. There's no custom pixel entry - deliberate, since H3's grid doesn't love arbitrary sizes. fps is locked to 24 and the range is 24–24, so it's a label, not a knob.

handoff_frames is how many frames each segment takes from the end of the previous one. The tooltip is admirably specific: the panel exposes 5 or 22, and anything you type from 1–5 quantizes to 5 while 6–22 quantizes to 22. More frames, more stable seam, more compute.

handoff_mode picks how that handoff travels. 尾帧续写（latent） passes the previous segment's tail latent straight through - lossless, fastest, the same trick the Motion Context path uses. 尾帧画面（RGB） re-encodes the tail as a visual guide: a tighter visual match, at the cost of one extra decode per segment.

steps / denoise / sampler_name / scheduler are plain sampling controls with euler + normal + 20 steps as the defaults. If you've connected DMAD's four-step accelerator, ignore these - see below.

exposure_lock (on by default) is the seam fixer: the first 12 frames of each segment match the previous segment's tail colour, a 12-frame additive brightness ramp smooths the join, whole-segment correction is capped at ±0.04, and a single-frame pose jump inside the same shot gets a light motion bridge while hard cuts are left alone. Turning it off gives you exactly what the sampler produced, warts included.

The settings that exist because big models are big

MiniMax H3 is a 33B omni-modal model with native audio, and the day-one "it fits on consumer hardware" claims drew instant scepticism. The weights are reportedly ~42.5GB. This node's job is to make that survivable:

• vram_tier - 自动（按显卡） reads your card once and picks. 16G is effectively "no extra limits" (the old behaviour). 12G caps the canvas at 720P, unloads the text encoder before sampling and the trunk before decoding, and forces preview interval ≥ 2. 8G caps at 480P, uses a lighter RGB preview and forces latent handoff. If you're getting OOM at 720P on a 12GB card, this is the first widget to check, not the sampler.
• vram_policy - per-segment cleanup: 释放缓存 (recommended) hands cached VRAM back to the driver but keeps the model resident, so the next segment starts fast. 卸载模型 frees the most but reloads weights every segment, which on a long film is a brutal tax. 不处理 is ComfyUI's default behaviour.
• sample_preview / preview_interval - live step previews during sampling. Lovely on a 24GB card, ruinous on an 8GB one, which is why the tier profile overrides it.

The ones people miss

audio_mode - 生成音频 gives every segment its own generated sound. 数字人（锁定音频） uses the single audio track in the media library as the driving track: sliced along the timeline, locked into each segment's latent so the picture follows the audio, and reused for the final film. No audio in the library and it quietly falls back to generated.

keep_full - in RGB-handoff mode, keeps the ~12 extra grid-alignment frames per segment so sentence endings and seams land better; the film runs slightly longer than the seconds you typed.

ref_image_size - how reference images are sized before encoding (1k default, match, 1.5k, 2k, original). 2k costs VRAM and helps faces; that's the whole trade. prompt_optimizer_settings opens the optimiser panel (API or local model, llama-cpp-python for the local route).

And the two optional inputs that override half the panel: connect a sampler (e.g. DMAD's re-noise sampler) and a sigmas (DMAD Sigmas) and they override sampler_name, scheduler, steps and the denoise truncation entirely. The pack's own release notes recommend DMAD at LoRA strength 1.0, 4 steps, shift 12 with the audio shifted separately. Unconnected, the built-in values apply as normal.

Inputs and outputs

Required h3_bundle, then the widgets above; optional model, sampler, sigmas. One output: sequence_config (MINIMAX_H3_SEQUENCE_CONFIG) into the first 视频段落 node. Later segments inherit it down the previous_segment chain, so you wire it once, not twenty times.

Install

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

Restart, Ctrl+F5, under the AICG3D category. Nothing extra for this node; DMAD is a separate optional pack, and the local prompt optimiser needs llama-cpp-python.

Gotchas

Changing resolution mid-workflow invalidates everything downstream - every segment resamples at a new canvas. Set it before you write the prompts, not after. And if a long chain dies at segment 14 with a CUDA OOM, don't touch the sampler settings: step down vram_tier first, then the resolution.