H3 Continuum Sampler V3.6

The one node that makes H3 feel like more than a 5-second toy

MiniMax H3 is genuinely good - 33B omni-modal, native stereo audio, day-zero ComfyUI support - but out of the box it gives you short clips that hard-cut at the chunk boundary. If you've run a few H3 workflows you know the pain: five seconds of glorious motion, then a hard cut to something that doesn't continue it. H3 Continuum Sampler V3.6 is the pack's current production answer: it generates long-form video as a chain of chunks while carrying the previous video and audio forward, so your 15-second piece actually behaves like one continuous take.

The idea is simple - generate chunk 1, then chunk 2 knowing chunk 1, and so on up to 16 chunks. Everything interesting is how well the next chunk knows the last.

How the continuation actually works

Older H3 continuation in this pack appended the prior frames as a separate "Reference Context" block alongside the new target. V3.6's Standard backend does something different: it places the finalized video and audio latent prefixes inside the next H3 target and protects them with Core noise masks. That's the "Masked AV" path - the protected prefix can't be resampled, so the prior chunk's ending survives bit-exact while the new region is generated around it.

This is both cheaper and better behaved - in the pack's gates, Standard removed about 8% of the packed rows versus Compatibility and shaved a few percent off median sampling time.

One catch trips people: Masked AV only activates for Balanced - 22 frames when Audio Continuity is on. Pick Fast - 5 frames, Strong - 39 frames, or Auto and Standard safely falls back to the older Reference Context transport. The status output tells you which transport resolved - read it.

The inputs that actually matter

Most defaults are sane. The ones you'll actually set:

• sequence_prompt - connect one multiline text node for the whole sequence. Use the Timeline format: each [0-5s]-style header on its own line, description on the following lines. Do not put the prompt on the same line as the header - it silently gets treated as a Fixed prompt and every chunk reuses it.
• chunks and chunk_seconds - 3 chunks × 5 seconds is the natural start. chunk_seconds accepts up to 30, but 5–15 is the validated range; above 15, VRAM and runtime climb fast.
• width/height - defaults 1344×768, keep multiples of 32.
• continuity - Balanced - 22 frames is the default and the one that unlocks Masked AV. Leave it there.
• run_storage + run_name + reroll_from_chunk - the killer feature. Save the run, then set reroll_from_chunk to "Chunk 2" and only the second chunk regenerates; the first part stays byte-identical.

Optional inputs worth knowing: first_frame/last_frame (leave disconnected for T2VA), reference_image_1–3 for persistent identity, reference_video_1 (Video Guide Frames, from a video loader at 24 fps), and driving_audio + audio_vae when an existing track should be the final audio.

Outputs wire like this: video_latents and audio_latents go to Core Video/VAE Decode nodes, assembly_plan (plus the decoded images/audio) into H3 Continuum Assemble + Seam, and refine_context into the pack's Second Pass or Hi-Res Fix nodes. status is a string - read it, it reports the resolved transport and chunk reuse.

Install

ComfyUI 0.32.0 or later. Easiest via ComfyUI Manager - search "H3 Continuum". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git

Restart after. The pack's requirements.txt is empty - no pip packages, just the PyTorch and safetensors ComfyUI already has. You still need the MiniMax H3 weights, which is where people actually get stuck: the H3 Community License excludes the US, EU, UK and South Korea from running the local weights at all, so check that before you go further.

Where people get burned

• Wrong prompt formatting is the #1 silent failure. Header on its own line, always.
• Audio in the wrong socket. For a video with sound, route the loader's IMAGE to reference_video_1 and its AUDIO to driving_audio. Don't feed that audio into reference_audio_1 unless you want conditioning-only behavior - that one does not replace the final audio.
• Video Guide Frames at the wrong frame rate. Feed it a 24 fps source or set force_rate: 24 on the loader.
• Long, high-res runs can OOM. The pack's own 16 GB GPU test died on a 3×5s 1152×1152 Hi-Res Fix run, so start at 576–640px for anything multi-chunk. If a run dies mid-way, that's what run_storage is for.