H3 Multishot Sampler (one node)

The classic one. MiniMax-H3 generates roughly 10-15 second blocks, and if you've ever tried to chain two of them with stock nodes you know what happens: shot 2 re-imagines the scene from text, so the face shifts, the wardrobe drifts, the colour jumps, and what you have is a cut sequence, not a take. H3MultishotSampler is the pack's zero-dependency answer to that - it renders shot after shot and hands each one's last frame to the next as its first frame, using the checkpoint's own trained continuation task (fl2va's image-to-video mechanism), so the join is continuous by construction. Same face, same room, same light - because the next shot starts from an actual rendered picture, not a fresh reading of the prompt.

How the chaining works

Each shot's last frame becomes the next shot's first frame through fl2va's trained I2V path - the conditioning route the model was trained on, not a bolted-on hack. Two cleanups make the join invisible: the duplicated boundary frame is trimmed (one-frame stutter gone), and the seam audio gets a 40ms equal-power weld so the join doesn't dip in loudness. Identity holds through a second mechanism as well: byte-identical text. The prompt writer repeats each character's appearance block verbatim in every shot, so sampling noise around a reworded description can't drift the face. Frame pins the instance; text pins the category.

The inputs that matter

• script - one prompt per shot, --- on its own line between them (JSON {"prompts": [...]} also accepted). This is where the quality lives. The README's boundary rules are worth reading before you write one: each shot after the first opens holding the previous shot's closing arrangement with ~2 quiet seconds, and a spoken line never straddles two shots.
• shot_count - 0 = one shot per --- block (leave it). 1-8 forces a total; a short script repeats its last block as a continuation.
• frames_per_shot - 243 default (~10.1s), 362 is the trained max (~15.1s), on H3's 17k+5 grid.
• seed_per_shot - leave ON. Measured: per-shot seeds hold the face; one seed for every shot drifted both face and voice.
• start_image - optional I2V first frame for shot 1 (a real first frame, unlike the memory sampler's same-named input). Later shots keep chaining from previous last frames.
• reference_images - character refs carried into every shot as <Picture 1>, etc. Bind them in the prompt: "She looks like the woman in <Picture 1>." Verified on ref2va.
• voice_ref - an audio anchor (<Audio 1>) that pins the voice across the chain. Also ref2va territory.
• self_anchor_voice - after shot 1 renders, its own audio becomes the reference for later shots. No file needed; write shot 1 so the character speaks a clean solo line.
• output_scale / upscale_model - post-decode resizing / real upscale. Read the tooltip carefully: with an upscale model wired, 1.0 does not mean off - it means don't correct, so you get the raw 4x. The console prints the projected size when the model loads; at 1344x768 a 4x upscale is 22GB a shot in system RAM.

Outputs: master_frames and master_audio (the joined take, seams trimmed), shots_rendered, and first_shot_frames / first_shot_audio so you can judge shot 1 before the chain commits.

Install and reality check

cd ComfyUI/custom_nodes
git clone https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

or search H3 Multishot in ComfyUI-Manager (Registry: comfyui-h3-multishot). ComfyUI v0.30.0+, no pip dependencies. Models: checkpoint in models/diffusion_models/, encoder + VAEs in text_encoders/ and vae/. GGUF users also install ComfyUI-GGUF.

The pack is young - effectively zero community footprint on Reddit as of writing - and its README is refreshingly honest about limits: audio dulls slightly per hop on long chains (restart chains on scene cuts), and the texture ratchet is ~1.05-1.3x per join depending on settings. For 3-4 shot scenes the joins reviewed as one continuous take. It's a meticulous solo project (jlucasmcrell / joeygambino), free, and the docs are the best part of it - read the prompting rules before you queue, and a word clipped at a join is almost always your script, not the sampler.