H3 Continuum Sampler V3.4

H3 Continuum Sampler V3.4 is the engine room of the H3 Continuum pack. You reach for it when plain MiniMax H3 in ComfyUI hands you a lovely ten-second clip and you want a minute of something that actually holds together - same scene, same character, audio that doesn't hard-cut at the end of every generation.

Set expectations before you clone anything: this is a small, quiet pack (these node pages see essentially zero traffic, and there's no real Reddit footprint to point you at), and it's an orchestration layer, not a model. It sits on top of ComfyUI's native MiniMax H3 Core nodes, which you still have to install and download. What the sampler adds is chunked continuation: instead of one H3 pass that dies at the model's clip length, it generates chunk after chunk, feeding the tail of each one into the next so the result reads as a single take rather than three clips glued together.

How it works

The sampler takes the Core H3 pieces - model, clip, video_vae, sampler, sigmas - and runs the model chunks times (default 3, up to 16), each of chunk_seconds (default 5, and 5 is the validated default; don't chase 15 on day one). At every boundary it carries over prior video context, which is exactly what the continuity control sets: Auto-conservative, Balanced at 22 frames, Fast at 5, or Strong at 39 frames (experimental). That carryover is the whole trick - it's why chunk 2 knows chunk 1 ended.

Your text goes in one sequence_prompt (a Text Multiline node) and prompt_mode in Auto figures out the style: Fixed reuses one prompt, List uses --- separators, and Timeline splits by [0-5s], [5-10s] blocks so you can direct each section. If it can't parse a timeline it warns and falls back instead of dying, which is the V3.4 "core-first, permissive" philosophy in action.

The inputs that matter

Most of the schema is comfortably advanced, but these are the ones you'll actually touch:

• first_frame / last_frame / reference_image_1–3 - the I2V and reference conditioning. Leave every image input disconnected for pure text-to-video.
• reference_video_1 - the persistent Video Reference. It's an IMAGE frame batch (loader outputs feed straight in) interpreted at 24 fps and applied to every chunk. video_reference_size keeps the token cost sane: Efficient 0.4 MP is the default, Match Output is much slower.
• driving_audio + audio_vae - the reason V3.4 exists. Your original audio is used as native H3 guide conditioning and selected unchanged for the final output, so a song or dialogue track survives chunking intact. audio_vae is only required when Driving Audio is connected.
• run_storage + run_name + reroll_from_chunk - Run Storage atomically saves raw AV chunks, so after a restart you can regenerate only the chunk that's wrong instead of the whole clip. Use a fixed base_seed (per-chunk seeds are derived deterministically from it) or resume becomes a crapshoot.

One gotcha the README gets wrong: it claims strict_compatibility was removed from the V3.4 interface, but the node still exposes it under advanced (default on). Leave it on.

What comes out

• video_latents and audio_latents (lists) - feed these to Core VAE Decode and Core VAE Decode Audio. Continuum never decodes; that's Core's job.
• assembly_plan - the boundary metadata that tells the Assemble node where seams are.
• status (string) and driving_audio (pass this to the assembler to keep the preserved audio).

Install and the real gotchas

Install via ComfyUI Manager (search "H3 Continuum" or just "Continuum"), or:

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

Restart ComfyUI. There are no extra pip dependencies - the requirements file is deliberately empty, using ComfyUI's bundled torch and safetensors - but you need ComfyUI 0.32.0+ for native H3 Core support, and the H3 model files come from Core, not this pack.

Two things will actually bite you. First, the V3.4 release shipped incomplete: if you installed V3.4 early, Driving Audio and Video Reference inputs didn't exist yet, so git pull (or Manager Update) and restart. Second, Video Reference really wants 24 fps source. Load Video (Upload) will accept a 25 fps file and quietly misalign it - set force_rate to 24 or convert first.