🎬 H3 Multishot - LD

Multishot is how you get past H3's fifteen-second wall

H3 samples a clip at a time - trained around 4–15 seconds, with a hard-ish ceiling beyond that. So a real scene is several clips that need to feel like one continuous take: same person, same voice, same room. The H3 Multishot Sampler exists for exactly that. It takes a ----separated script, samples each shot in sequence, and chains each one from the previous shot's last frame so the action carries across the boundary.

H3 Multishot - LD (🎬 H3 Multishot - LD, under LD / PromptMaster / Long form) is that upstream sampler with one important upgrade: H3 Studio's latent can drive it. Connect Studio's latent output and the node reads width, height and frames-per-shot from its shape - the FRAME bar in Studio becomes the only place size is decided, so the writer and the sampler physically cannot disagree. Leave latent unconnected and it behaves exactly like the original upstream node.

The two dials that decide whether your characters survive

• script - one prompt per shot, --- on its own line between them. JSON {"prompts": [...]} also works.
• shot_count - the total number of shots, not shots-per-prompt. 0 (the default) lets the script decide: one shot per --- block, which is what you want for a written scene. Force 1–8 and extra blocks get dropped.
• seed_per_shot - leave it ON. This is the one the docs measured, and the result is counterintuitive: varying the seed per shot holds the face across the chain, while one seed for every shot made both the face and the voice drift. Identity lives in the conditioning, not the seed.

Then the reference slots: start_image seeds shot 1 for image-to-video; reference_images are character references carried into every shot as <Picture 1> etc.; voice_ref is a clean solo line of your character fed as <Audio 1> so the voice is pinned instead of re-performed each shot. If you don't have a clean vocal line, self_anchor_voice does it automatically - after shot 1 renders, its own audio becomes the reference for every later shot, no file needed. Write shot 1 so the character actually speaks.

The finishing stack

Two knobs do the quality work after decode: upscale_model (wire ComfyUI's Load Upscale Model - ESRGAN and friends - to synthesise real detail per shot) with output_scale as the final size multiplier, and transition (cut, or a fade/dissolve applied to rendered frames so it's exact - audio is ducked to match) with transition_frames for how long each side lasts. save_every_shot writes each shot to output/video/H3_SHOTS/ the moment it decodes - insurance against a mux OOM at the end of a long chain destroying everything. vram_guard evicts the DiT before each text encode; off by default because it costs a model swap per shot, but flip it on if you hit Fault failed: 2 on a card that doesn't fit.

You can also hand the sampler a SAMPLER node (the MiniMax H3 Turbo sampler included) or a SIGMAS schedule to replace its dropdowns, which is how a turbo or LoRA curve stays consistent across the whole chain.

Outputs and install

Outputs: master_frames (IMAGE), master_audio (AUDIO), and shots_rendered (INT) - the concatenated deliverable plus a count.

cd ComfyUI/custom_nodes
git clone https://github.com/Lora-Daddy/PromptMasterLD
cd PromptMasterLD
pip install -r requirements.txt

Restart (or ComfyUI Manager → "PromptMasterLD"). This node wraps the vendored upstream ComfyUI-H3-Multishot, which the pack loads under the hood - install the pack and the sampler comes with it. The place people go wrong is almost always conditioning, not the sampler: feed it reference images of a different face than your first shot and watch the identity fight itself across the chain. Keep one face, one voice, one room, and let seed_per_shot do its counterintuitive thing.