H3 Continuum Sampler V3.5

This is the node everything else in the pack hangs off. H3 Continuum Sampler V3.5 is how you generate long-form MiniMax H3 video in ComfyUI: instead of one bounded pass, it splits the run into sequential chunks - 5 seconds per chunk is the validated default - carries visual and audio context across each boundary, and lets you keep a supplied video or audio track in the loop. If you've been frustrated that single-pass H3 tops out around fifteen seconds, this is the workaround that actually holds continuity.

MiniMax H3 is the 33B omni-modal model that generates video with native stereo audio, and it got day-zero ComfyUI support - but its native single-pass path is short. Continuum's bet is chunking with preserved context: at every boundary it carries a "continuity" window of prior video (the default is Balanced - 22 frames; Auto is conservative, Fast trims to 5, Strong's 39 frames is marked Experimental), so the next chunk starts from where the last one ended rather than from scratch. For 5-second FL2VA runs it also does what it calls Terminal Merge: the final two logical chunks become one physical sampling-and-decode unit so Last Frame handling matches Core behavior. Outputs come back per physical group, which is why several of the V3.5 nodes return lists.

The inputs a beginner actually sets

The sampler has a lot of widgets. Here's what matters:

• model, clip, video_vae, sampler, sigmas - the standard H3 stack. The VAE is used only to encode image conditioning; Continuum never decodes with it.
• sequence_prompt - one Text (Multiline) for the whole run, plus prompt_mode (Auto / Fixed / List / Timeline).
• chunks (default 3) and chunk_seconds (default 5) - your total length is roughly chunks × chunk_seconds.
• width / height - multiples of 32; defaults are 1344×768.
• continuity, base_seed, audio_continuity (on passes prior audio context into continuation chunks).
• reroll_from_chunk + run_storage (Save + Auto Resume) - for reusing completed chunks and regenerating only the part you don't like.

Everything else is optional connectivity: first_frame, last_frame, reference_image_1 through reference_image_3 (persistent identity refs), reference_video_1 (that's the Video Guide Frames socket - feed it a video loader's IMAGE frame batch, not a still), driving_audio plus audio_vae (your audio track, preserved as the final audio), and reference_audio_1 plus reference_audio_vae (H3 conditioning only, does not replace the output audio). Leave all image/audio inputs disconnected and you're doing text-to-video.

Outputs

video_latents and audio_latents (lists), assembly_plan (the map everything downstream reads), status, driving_audio (your preserved track, if connected), and - new in V3.5 - refine_context, which is what makes Hi-Res Fix and Second Pass context-aware. Wire the first three into Core VAE Decode + Assemble + Seam, or the last one into the refine nodes.

Installing it

ComfyUI Manager, search "H3 Continuum", or:

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

Restart after install. No pip deps beyond ComfyUI's bundled torch. The H3 weights (~42.5 GB) install the normal way, and be aware of the MiniMax H3 Community License's territory restrictions. For First/Last + Reference in one run you want a hybrid-capable model, e.g. the B2049 variant the README links.

Where people get burned

• Timeline prompts: the time header must be on its own line. [0-5s] then the description on the next line - [0-5s] Describe the scene isn't recognized and Auto falls back to Fixed, silently reusing one prompt for every chunk. Text before the first header becomes a global preamble applied to all chunks.
• Don't rely on "continue from the previous chunk." Continuum carries context, but describing the next action explicitly gives H3 a much better shot at not drifting or cutting.
• Video Guide Frames is a guide, not a copy, and it expects 24 fps input (set force_rate to 24 on the loader). It won't reproduce every frame.
• Driving Audio vs Reference Audio: driving audio replaces the final output's audio; reference audio only conditions it. Connect the wrong one and you'll wonder why your generated sound is still there (or isn't).
• Run Storage needs a fixed seed for reproducible resume, and changing models, references, prompts, or resolution creates a new revision, so earlier chunks stop being reusable.
• No universal memory guarantee: the README notes two-chunk 800×800 runs pass on a 16 GB card, but long high-resolution sequences can still exceed system RAM during decode and assembly.