[Legacy] H3 Continuum Sampler V3

H3 Continuum Sampler V3 was the big architectural reset for this pack: sampling, native continuation, State/Session handling, and Spectrum interop all stay inside Continuum, but video and audio VAE decoding moved out to ComfyUI Core. The V3.3 and V3.2.4 samplers you see today are direct descendants of this node. It's marked "[Legacy]" now - the pack keeps it registered for saved workflows - but if you want to understand how the modern ones work, this is where the current design started.

Before V3, the pack decoded and re-encoded audio/video itself between continuation segments. V3 flipped that: it samples all N chunks in latent space, carrying raw video/audio context between them, then hands you lists of latents plus an assembly_plan so you (or Core nodes) do the VAE work. That's why the README's "why Core decode is external" section exists - leaving decode to Core means tiled decode, memory strategy, and future H3 VAE improvements come to you with ComfyUI updates instead of waiting on the pack.

The inputs that matter

Standard pipeline: model, clip, video_vae (image conditioning only - V3 never decodes with it), sampler, sigmas, and sequence_prompt (Text Multiline) with prompt_mode = Auto. chunks (3, max 16), chunk_seconds (5.0), width/height (1344×768 default, multiples of 32), continuity (Balanced - 22 frames), and base_seed.

Optional: first_frame (leave all image inputs disconnected for pure T2VA), prompt_overrides from the Clip Overrides node, and advanced from H3 Continuum Advanced V3 - the bundle node that carries audio_continuity, reroll_from_chunk, reroll_nonce, diagnostics, and session/state inputs so the sampler surface stays clean.

Outputs: video_latents and audio_latents (lists, one per chunk → Core VAE Decode / VAE Decode Audio), assembly_plan (→ a Continuum Assemble node), and result (an H3_CONTINUUM_RESULT you can unpack with the Result node to get the last state, session, and a report).

Install

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

or install "H3 Continuum" via ComfyUI Manager, then restart. No extra pip dependencies - just a working MiniMax H3 ComfyUI environment (weights not bundled) and ComfyUI v0.32.0+.

Where people get burned with the V3 generation specifically: forgetting that this node never decodes. If you connect only result to something and see no video, you skipped the Core VAE Decode step that V3 delegates to. Also note this is the version before Run Storage - resume/reroll on V3 relies on the Advanced V3 node's session/initial_state inputs and the save/load state nodes, not the automatic disk-backed Save + Auto Resume that arrived in V3.2.4. For a fresh project, use H3 Continuum Sampler V3.3; keep this node around for the workflow files that still reference it.