H3 Auto Prompt Chain v2

MiniMax H3 tops out at 15 seconds a clip. That's great for a b-roll loop and useless for a scene, let alone a short film. This node is the thing that closes that gap: you give it a list of shot prompts, and it sits there generating shot after shot, continuing each new clip from the latents of the last one, and finally stitches everything into a single video. It's called "H3 Auto Prompt Chain v2" in the node menu, and it's the closest thing to a one-click long-form pipeline the H3 ecosystem has right now.

Quick context if you're new to H3: it's MiniMax's 33B omni-modal video model - native stereo audio, up to 2K, released open-weights in August 2026 with day-zero ComfyUI support. It's also geofenced by its community license: if you're in the US, EU, UK, or South Korea, you're not licensed to run the local weights at all, so read the license before you download. Everything below assumes you're somewhere that can legally run it.

What it actually does

This is an output node - you wire up your model plumbing and it runs when you hit Queue, unattended, up to 80 shots in one pass. Under the hood it's an orchestrator for the Herrgotts-H3-Infinite-Continuation-Suite: shot one is a normal text-to-video conditioning call, then every subsequent shot is built through the suite's H3ContinuousContinueV11 node with a phase-aligned handover. Crucially, shots are chained in latent space - no VAE decode/re-encode between clips, which is where a naive chain loses both time and coherence. The sampler sigmas are cached per step count so the whole run isn't rebuilding them, and there's a single memory-bounded stitch/decode at the very end. The author's docstring recommends feeding it a turbo-style H3 sampler from that suite for speed; a stock sampler works too.

The inputs that matter

You need the full H3 plumbing - model, clip, video_vae, audio_vae, and a sampler - straight from the standard H3 loaders. After that, the one you'll actually live in is shot_prompts. One shot per line, or separate multiline prompts with a line containing ---. Each shot can carry leading tags that override the run's defaults:

• [FAST] → 4 steps, 5 context frames · [BALANCED] → 6/22 · [QUALITY] → 8/22 · [MOTION] → ≥6 steps, 22 context
• [dur=10] (5–15s), [steps=6] (1–40), [ctx=22] (5/22/39), [tail=7] (0–68 landing frames)

global_prompt carries the continuity instruction ("preserve character identity, wardrobe, lighting…") - the default is a solid starting point. start_clip is your resume knob: if clip N dies mid-run after earlier latents were saved, set it to N, keep the same prompt list, seed, and latent prefix, and it picks up from the saved clip instead of redoing everything. base_seed + shot index drives per-shot seeds, and the defaults are low-VRAM friendly (608×352, 6 steps). If you want to anchor the very first shot, the optional first_frame image input does exactly that.

Outputs

Three strings: video_path (the stitched final file), run_info (a completion summary with timing), and profile_path (a JSON with per-shot conditioning vs. sampling seconds - genuinely useful for finding your bottleneck, since H3's prompt encoder can cost more than the diffusion). Everything lands inside ComfyUI's configured output directory.

Installing it

ComfyUI Manager: search "H3-AutoPromptChain" and install. Or:

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

Then restart ComfyUI. No extra Python packages - but you must have the Herrgotts suite installed too (clone it the same way), because it provides the continuation nodes. Skip it and you'll get a clear RuntimeError naming H3ContinuousContinueV11. That's actually the friendliest failure mode this pack has.

Where people get burned

• Forgetting the Herrgotts suite. The pack doesn't bundle it and won't install it for you. Error message tells you exactly what's missing.
• Resuming wrong. When you resume from start_clip N, you must keep the same prompt list, seed, and latent_prefix - change any of them and the "continuation" is continuing something else.
• Dimensions and caps. Width/height step in multiples of 32, and 80 shots is the hard ceiling per run.
• The license. Not a bug, but the geofence is real, and it's the kind of thing people discover embarrassingly late.

If you're generating H3 content longer than a single clip, this is the one to reach for. It's young and it's a thin wrapper around a specific continuation suite - but "paste a prompt list, wake up to a stitched film" is exactly what's missing elsewhere.