🎬 H3 Master - LD

The node that turns Studio's script into an actual movie

H3 Studio writes you a script. H3 Master (🎬 H3 Master - LD, under LD / PromptMaster / Long form, in the same PromptMasterLD pack) renders it: one ----separated script in, one continuous IMAGE+AUDIO video out, in a single queue. Wire Studio's pack output into pack and Master handles everything between - the pack carries the script, seconds, canvas and rails, so wiring it alone gives you a clean chain. (The script textarea is there for when you're writing prompts by hand instead, and its --- format is the same contract Studio's writer already emits, so scripts port unchanged.)

Why does a whole movie need one node instead of a graph? Because clip N+1 needs clip N's decoded pixels, and ComfyUI evaluates a static graph exactly once - a loop can't be drawn. So the loop lives inside the node's run(). And rather than reimplementing H3's chaining, Master calls the community's proven machinery directly: NikoDemon80's MiniMax H3 Motion Context node pins the previous clip's tail as interior keyframes on the next clip's timeline (plus its audio), then trims the duplicated head on both picture and sound. Measured on real renders, joins land at roughly one ordinary frame step of picture and 0.00ms of audio drift. It works because it stopped reinventing the thing that already worked.

The dials that matter

• script - one prompt per clip, --- between them. This is the same contract H3 Studio emits, so scripts port verbatim.
• frames_per_clip - snaps to H3's 17k+5 grid (a frame count that lands on a valid H3 length). 328 is about 13.7s; clips after the first render this minus the carry.
• carry_frames - frames of the previous clip pinned into this one, the last ~1s at the seam. Official grid 5/22/39/56; type 1 and it becomes 22.
• multishot mode - with MULTISHOT ticked in Studio, Master renders N clips and joins them rather than a single long shot, which is the sane way to get past H3's ~15s per-clip ceiling.
• seed, width, height - the usual, plus a seed_per_clip option to vary seeds per clip.

The optional inputs are where the long-form life gets good. extend_video / extend_audio / extend_latent continue an existing clip instead of starting from black - wire a video loader's frames and the chain starts from that clip's tail, the last carry_frames becoming clip 1's motion context. Connect the matching .safetensors latent and the join is bit-exact. There's also declick_ms (default 1ms - ramps both sides of each join through zero so you don't get an audible tick at every cut), a cache keyed on (seed, index, prompt, geometry) so editing clip 3 re-renders only clip 3 onward, and the music_mp3/music_plan pair if Studio wrote a music-video script.

What comes out

Four outputs: frames (IMAGE), audio (AUDIO), clips (INT, how many clips were rendered and joined), and report (STRING). The frames+audio pair is video-ready - feed it to a save node and you're done. If the joins still feel abrupt, seam_blend adds a few frames of dissolve across each cut using the head frames the trim would otherwise throw away - those frames are a second rendering of the same moment, so the dissolve is real overlap, not a fade to black.

Install, and the honest traps

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

Restart (or ComfyUI Manager → "PromptMasterLD"). Master relies on the Motion Context node, which the pack vendors under vendor/ComfyUI-H3-Motion-Context - but that upstream pack is GPL-licensed, and bundling it makes this distribution GPL as a whole. If you need permissive, the README tells you to delete that vendor folder, remove it from _VENDORED in __init__.py, and install it separately. H3 Master raises a clear error when it's missing, so the failure mode is loud, not silent.

Two practical traps. Extending is the place people lose a day: without extend_audio, a join is silent on the audio side, because there's no previous clip to carry sound. And remember the licensing map - H3's weights themselves are community-licensed and geofenced out of the US, EU, UK and Korea, so check before you put local weights in production in those regions.