[Legacy] H3 Continuum Sampler

H3 Continuum Sampler is the compact facade over this pack's legacy V2 execution core - one node that runs an entire N-chunk MiniMax H3 continuation: prompt/image conditioning precomputed, raw video/audio latent context carried between chunks, and decoded images plus audio handed back at the end. The description is blunt about why it exists: it's the V2 core "without dynamic sockets or frontend display patches." Meaning the fancier V2 node got the pretty UI, and this one just does the work with a fixed socket layout.

It's the node to know if you've opened a saved workflow and found a single unassuming sampler doing a surprising amount of work. The inputs are the full V2 sampler set minus the session plumbing: model, clip, video_vae, audio_vae (yes, V2 decodes internally, so the audio VAE is a real input here), sampler, sigmas, sequence_prompt, prompt_mode (Auto), chunks, chunk_seconds, width/height, continuity, base_seed, and seam_correction (Auto - native-first, only adopting video correction when the boundary and five-frame transition clearly improve).

The three optional inputs are where the compact facade hooks into the rest of the ecosystem: first_frame for I2VA, prompt_overrides from the Clip Overrides node (per-chunk prompt sections without a wall of sockets), and advanced from the H3 Continuum Advanced node - which is where audio_continuity, reroll_from_chunk, reroll_nonce, strict_compatibility, exact_total_duration, and diagnostics live. That's the design: the sampler stays small, and you attach the settings bundle only when you need it.

Outputs: images, audio (both ready for save/Create Video), and result (H3_CONTINUUM_RESULT) - unpack the latter with the Result node to get last_state, session, and the run report.

Install

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

or via ComfyUI Manager ("H3 Continuum"), then restart. No extra pip packages; needs the MiniMax H3 ComfyUI environment (weights not bundled) and ComfyUI v0.32.0+.

For new projects, don't reach for this one - H3 Continuum Sampler V3.3 is the current node, with Run Storage auto-resume and external Core decode. But if a shared workflow or an older tutorial drops this in front of you, everything here still runs as-is. The one thing that trips people: result is not video, so don't try to preview it directly - that's what the Result node's report and last_state outputs are for.