[Legacy] H3 Continuum Sampler V2

H3 Continuum Sampler V2 is the do-everything version of this pack's sampler, and understanding it explains why the later versions look so different. It precomputes prompt and image conditioning, isolates accelerator runtime state per chunk, captures latent state before decoding, then decodes and assembles the final AV stream itself. One node, whole video out. It's marked "[Legacy]" and the pack now recommends the V3/V3.3 samplers - but V2 workflows are still everywhere, and this is the node that made the pack's name.

The V2 output is friendly in the best way: images and audio come out already decoded and assembled, ready for a save or Create Video node. You also get last_state and session outputs (for continuing or persisting the run) and a report STRING. No separate VAE decode step, no assembly node - that all happens inside. The tradeoff is that the pack owns the decode, so it doesn't ride along with ComfyUI Core's H3 VAE improvements the way V3+ does.

The inputs that matter

Core pipeline: model, clip, video_vae, audio_vae, sampler, sigmas. Note the audio_vae input - V2 decodes audio itself, unlike V3 which leaves that to Core. prompt_mode (Auto) plus prompt_script (Text Multiline) parse Fixed / --- List / [0-5s] Timeline prompts, and sequence_prompt is an optional socket so you can drive the script from elsewhere. chunks (3, max 16), chunk_seconds (5.0), width/height (multiples of 32), continuity (default Balanced - 22 frames; Auto - conservative scores latent motion and falls back to 22f), base_seed, and audio_continuity (true).

The session/reroll system lives here too: reroll_from_chunk (0) reuses matching accepted chunks from an optional session and regenerates N onward; reroll_nonce changes seeds for regenerated chunks only; strict_compatibility (true) runs proactive H3 API checks. There's also seam_correction (Auto) - native-first video correction adopted only when the boundary and five-frame transition clearly improve, with audio guarded independently.

V2 also carries up to sixteen clip_1_prompt..clip_16_prompt override sockets. That's the old way to give each chunk its own prompt; the newer Clip Overrides node replaces that pile of sockets with one prompt_overrides input.

Optional inputs: first_frame, last_frame, session, initial_state, prompt_plan (from the Prompt Plan Preview node), and show_preview.

Install and when to care

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 a MiniMax H3 ComfyUI environment (weights not bundled) and ComfyUI v0.32.0+.

If you're loading an older V2 workflow, this node is why it still works - the pack explicitly keeps legacy IDs registered for compatibility. If you're building fresh, skip it: the V3.3 sampler produces better-behaved output (external Core decode, Run Storage auto-resume, guarded seam correction on the assemble side) and fewer moving parts. Keep V2 in mind mainly for the handful of workflows that lean on its integrated decode or its session model.