H3-AutoPromptChain

H3 Auto Prompt Chain

A ComfyUI output node for unattended MiniMax H3 long-video generation from a list of prompts. It continues directly from saved audiovisual latents, supports per-shot generation budgets, can resume interrupted chains, and performs one final stitched decode.

Requirements

• A current ComfyUI installation with MiniMax H3 support
• Herrgotts-H3-Infinite-Continuation-Suite
• A compatible MODEL, CLIP, video VAE, audio VAE, and SAMPLER

No additional Python packages are required.

Installation

Clone the repository into ComfyUI/custom_nodes, then restart ComfyUI:

cd ComfyUI/custom_nodes
git clone https://github.com/misutesu-desu/H3-AutoPromptChain.git

The node appears as H3 Auto Prompt Chain v2 under MiniMax H3/automation.

Shot prompts

Put one prompt on each line, or separate multiline prompts with a line containing ---.

[BALANCED] SHOT 01: A woman walks toward the apartment window.
---
[FAST] SHOT 02: She looks down at the street.
---
[MOTION][dur=10][tail=7] SHOT 03: She turns and runs toward the table.

Leading tags can override the defaults for each shot:

• [FAST]: 4 steps, 5 context frames
• [BALANCED] or [NORMAL]: 6 steps, 22 context frames
• [QUALITY] or [HQ]: 8 steps, 22 context frames
• [MOTION] or [ACTION]: at least 6 steps, 22 context frames
• [dur=8]: duration from 5 to 15 seconds
• [steps=6]: 1 to 40 sampling steps
• [ctx=22]: 5, 22, or 39 context frames
• [tail=7]: 0 to 68 landing-tail frames

For a fresh run, leave start_clip at 1. If clip N fails after earlier latents were saved, set start_clip to N and keep the same prompt list, seed, and latent prefix.

Outputs

• video_path: final stitched video path
• run_info: timing and completion summary
• profile_path: JSON profile with per-shot conditioning and sampling times

Generated latents, videos, and profiles stay inside ComfyUI's configured output directory.

Notes

• Up to 80 shots are accepted in one unattended run.
• The final clip is saved without trimming its ending.