Silver H3 Staged Multishot (encode once / sample once / decode once)

Every open video model falls apart at the same place: the join. You make five good ten-second clips and then discover the character's jacket changed, the room shifted two stops, and the voice is a different person. SilverH3StagedMultishot is the pack's attempt to make that seam invisible - unlike the batch node next to it, it isn't pretending independent clips are a long video. It chains them.

The output is one continuous take, N shots long. Same three-phase discipline as Silver H3 Batch (all prompts encoded in one text-encoder session, all shots sampled with the DiT resident, one streamed decode pass), but the shots are wired to each other in latent space, so motion, exposure and speech carry through a cut instead of restarting.

How the shots join

Two modes, selected by the ref2v toggle rather than by which inputs you happened to connect.

With ref2v OFF (fl2v), start_image seeds shot 1 the way stock image-to-video does, and shots 2+ chain on the previous shot's actual last frame through H3's native first/last keyframe channel. Use an fl2v-capable checkpoint, and expect character refs to be dropped with a warning: fl2v keyframes and identity refs are contradictory by construction, because the ref latents replace the keyframe latents.

With ref2v ON, start_image is ignored and shots 2+ inject the previous shot's last frame as the first reference block - the next shot gets to re-angle around that anchor instead of being pinned to it. This is the mode where char_refs_all, char_refs_index and char_audio do their work.

Either way, motion is carried by a per-step latent handoff, and that's a slightly expensive trick: each finished shot's tail has to be decoded to build the next shot's reference, forcing a DiT ↔ VAE swap at every join. The text encoder still loads only once, and it's the biggest of the three models, so the staging wins - but the README is honest that the per-join swap is unavoidable, not a bug.

What to actually set

prompts (one per shot, a STRING_LIST - feed it from Silver H3 Line Batch or Silver H3 Prompt List), overall_prompt for the scene-wide master, and then width / height / frames_per_shot (243 default ≈ 10.1s on the 17k+5 grid).

The three continuity knobs are where this node's personality lives:

• latent_handoff_taper (15) is the percentage of a shot that carries the previous shot's motion as a linear fade. 15% of a 10s clip is 1.5s of carry-over. It makes a seam continue; it does not create a transition. If your prompt opens with its own cut, drop this toward 0–5% so stale motion stops fighting the description.
• audio_lock (ON) makes the next shot's audio head a locked replay of the previous tail, trimmed symmetrically on decode. Off means a silent join: the new head is kept whole and the previous tail is trimmed with a fade-out, so script each shot to land its line and hold still.
• audio_onset_guard (ON) parks the quietest ~0.4s of the previous shot's audio latent just past the locked span so the model can't plan speech underneath the replay.

context_frames (5, up to 39) is how many decoded tail frames feed the next shot as temporal context - 1 is stock single-keyframe behaviour, higher values buy continuity at the cost of a VAE encode per join. pin_renorm and color_reanchor are both on by default and both exist to stop brightness and white balance drifting down a long chain. guide_audio is the spine: encode the whole voice recording once and every shot's audio locks to its slice, which is the closest thing to "the voice cannot change" you'll get. double_pass and upscale_factor are the same refine-and-upscale pair as in the batch node.

Outputs are just master_audio and saved_video. The take lands in a dated subfolder as one mp4 per shot plus _master.mp4, each shot written and freed the moment it decodes - which is why a long run that dies on shot 7 leaves shots 1–6 on disk instead of nothing. Use those files: the mp4s are the trimmed, correct output.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/SilverAndJade/comfyui-silver-nodes

Restart, then either load SilverH3Staged_Turbo_3x10s.json from the repo or search "Silver Nodes" in Manager. The nodes use comfy_api.latest video types and PyAV, so keep ComfyUI current, and you need H3 weights - about 42.5GB, under the MiniMax H3 Community License, which excludes the US, EU, UK and South Korea. The repo's H3BATCH.md is the long-form reference and is genuinely worth reading.

Where it bites

A bad shot means re-running the chain. That's the price of coupling: shots are deterministic start to finish, but not independently re-rollable. If that matters more to you than continuity, use Silver H3 Batch, where the master is a hard-cut stitch rather than a transition.

Reusing an image batch shorter than your shot count won't gap-fill. char_refs_index reuses its last image for the rest, and there's no way to attach a reference to just shot 3. One voice at a time, too - no two-hander dialogue. The README's "not supported yet" section is accurate.

Updates can leave stale graphs, and expect [SilverH3][dbg] console noise on long runs - the forensics are compiled in at the moment.