[Legacy] H3 Continuum Sampler V3.2.4

Read the "[Legacy]" tag, take a breath, and don't panic: H3 Continuum Sampler V3.2.4 is the production sampler that most existing H3 Continuum workflows still run, and it works fine. It's the version that shipped in the V3.2.4 release, and the current V3.3 pack keeps it registered purely so saved workflows keep loading. If you're starting fresh, grab H3 Continuum Sampler V3.3 instead - it's the same engine plus Timeline Video and the integrated seam surface. But if your workflow JSON says V3.2.4, this node is the reason it still runs.

Functionally it's the full production sampler: 1 to 16 chained H3 chunks with raw video/audio latent context carried between them, first/last-frame and up-to-three-image Reference conditioning, reference audio, and the disk-backed Run Storage for resume and reroll. Sampling stays inside Continuum; VAE decoding happens in ComfyUI Core, so you wire the list outputs into normal VAE Decode and VAE Decode Audio nodes.

Inputs that matter

The standard H3 pipeline connections: model, clip, video_vae (image conditioning only - Continuum never decodes with it), sampler, sigmas, and sequence_prompt from a Text (Multiline) node with prompt_mode = Auto (detects Fixed, ----separated List, or [0-5s] Timeline prompts). chunks (3) and chunk_seconds (5.0) set the run length; width/height default to 1344×768 and stay multiples of 32; continuity defaults to Balanced - 22 frames. base_seed derives deterministic per-chunk seeds.

The interesting V3.2.4 additions over the older V3 sampler:

• run_storage (Save + Auto Resume) - atomically saves each completed raw AV chunk. Interrupted runs resume from disk on the next Queue, and compatible prior runs are reused. run_name and project_id optionally override the auto-derived storage identity; reroll_from_chunk (Auto) reuses the longest compatible saved prefix and regenerates from a chosen chunk; reroll_nonce only changes the seed for an explicitly regenerated chunk.
• reference_image_1..3 - Reference conditioning, compacted in Core connection order with bypassed sockets ignored. reference_size = Match Output (practical default) or Max Identity.
• reference_audio_1 + reference_audio_vae - one native H3 reference audio track across all chunks. Connect the Audio VAE only when reference audio is used.
• audio_continuity (true) - carries prior audio context forward; turn off only to isolate or replace the generated audio.

Outputs are video_latents and audio_latents (lists, one per chunk → Core VAE Decode), assembly_plan (→ H3 Continuum Assemble V3.2.4), and status (STRING → Core Preview as Text).

Install and gotchas

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 deps; you need the MiniMax H3 MODEL/CLIP/VAE environment (weights not bundled) and ComfyUI v0.32.0+. Turbo 8-step weights come from the LightX2V MiniMax H3 Turbo HF repo.

Two V3.2.4-specific notes. It predates the V3.3 timeline_video input and the guarded video_seam correction - the assembler side still has Audio Seam Auto, but there's no video seam control. And don't let the "Legacy" label talk you out of updating the pack: the V3.3 release explicitly says these node IDs are preserved for compatibility, so updating is safe and your V3.2.4 workflows keep loading. For new work, just use V3.3.