H3 Continuum Sampler V3.8

MiniMax H3 is a 33B omni-modal video model that generates its audio in the same pass as the picture - genuinely impressive stuff, and the first serious open answer to "video that comes with sound." But like most diffusion video, a single native generation tops out around 15 seconds, and if you've ever faked a longer clip by queuing three separate generations, you know how that ends: the camera drifts, the jacket changes color, the music shifts key at every cut. H3 Continuum Sampler V3.8 is the node that solves that properly. It's the "production sampler" of the H3 Continuum pack - it splits a long run into chunks, carries video and audio context across every boundary, and lets you review, keep, retry, or resume each chunk instead of gambling on one whole run.

An orchestrator, not a sampler

The name is a bit of a lie in one direction: it's not a sampler in the ComfyUI sense of "pick Euler vs. DPM++." It's an orchestrator that sits between your model and your decoder. You still give it a real sampler and sigmas, and it uses them unchanged for every chunk - what it adds is the chunking, the continuation, and the bookkeeping.

What goes in

The required graph inputs read like a Core H3 setup: model, clip, video_vae, sampler, sigmas, and sequence_prompt (one multiline text node with your whole sequence). One quirk: video_vae only encodes image conditioning - Continuum never decodes with it, and pure text-to-video doesn't touch it. Decoding happens downstream.

Then the widgets that actually run the show:

• prompt_mode - Auto detects whether your sequence_prompt is a Fixed prompt, a ----separated List, or a [0-5s]-style Timeline. Timeline is where long-form scenes live - put a global preamble before the first [0-5s] header, then describe each chunk's action in its own section.
• chunks and chunk_seconds - the final total (1–16) and each chunk's duration. 5–15 seconds is the validated sweet spot; 30-second chunks exist but eat VRAM at high resolution. Total Length is just their product - it is not a progress counter.
• size_source - First Image sizes the output from your first_frame's aspect ratio (with a preset from Draft - 0.30 MP up to Native 768); Manual uses exact width/height in multiples of 32 - the normal choice for text-to-video.
• continuity and audio_continuity - how much prior context survives each boundary. Balanced - 22 frames is the accepted default; Strong - 39 frames is experimental.
• base_seed - keep it fixed across queues if you want resume and regeneration to actually reuse your saved chunks.
• run_storage (Save + Auto Resume) with an optional run_name, plus reroll_from_chunk to rebuild from a chosen chunk, and generation_mode (Full Run vs Review Each Chunk) for human-in-the-loop work.

Optional media sockets round it out: first_frame/last_frame for I2VA/FL2VA, three reference_image_* sockets, reference_video_1 (the "Video Guide Frames" batch), driving_audio + audio_vae to preserve a source track, and reference_audio_*/audio_references for conditioning-only sound.

What comes out

Six outputs, but the pipeline you'll actually wire is the first two: video_latents and audio_latents (lists, one entry per chunk) go to Core Video/Audio VAE Decode, then assembly_plan tells Continuum's Finalize node how to stitch the decoded groups back into the exact requested timeline. status is a plain-text report of what the last queue did - read it before blaming the node for a mystery. driving_audio passes your selected source track through, and refine_context is only for the advanced Second Pass bridge.

Installing

It's the one-clone whole-pack install:

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

Restart ComfyUI, or search "H3 Continuum" in ComfyUI Manager. No pip dependencies - it runs on ComfyUI's bundled PyTorch and needs ComfyUI 0.32.0+ (verified against 0.34.2). The supplied example workflow also wants Spectrum, rgthree, KJNodes and ComfyUI-Easy-Use installed, but those are only for opening that one graph - the node itself doesn't need them.

Where people get burned

First: in Review Each Chunk mode the first queue intentionally produces one chunk - a 5-second video when Total Length says 30. That's the design; approve it and queue again. Second: the review action buttons only select what the next queue does - you still have to press ComfyUI's top-right Run. Third: V3.8 exports exactly seven nodes, so a saved workflow from an older release can load as an unknown node; use the matching release tag rather than fighting it. Two honest caveats from the pack's docs: long repeated continuation can drift contrast and sharpness (open issue #13, no production fix), and the accepted 16 GB gate ran near 15.5 GiB in the hardest cases - tested, not guaranteed. And if you're in the US, EU, UK or South Korea, double-check the MiniMax H3 Community License before building anything on the weights this node orchestrates.