A shot writer and pipeline pack for ComfyUI, built around H3 video. It runs a local model through LM Studio, turns a one-line intent into a finished shot script, and hands the graph the text plus real timing — then ships the sampler, routing, reference and audio nodes that render it.

The centre of the pack is the H3 Studio node and its panel. Everything else exists to get a script out of it and into a render.

What it does

You type what you want. The panel builds a brief from the dials you actually set, sends it to your local model, and returns a script the graph can use — with the frame count, the beats and the pack-out already correct.

Eight modes, chosen on the node:

| Mode | What it writes for | |---|---| | ref | reference-driven shots — the default; files you attach are cited by name | | i2v | image to video, opening on the frame you supply | | t2v | text to video, building the frame from nothing | | mv | music videos — per-shot start stills, a tagged cast rail, wardrobe lock | | multishot | multi-shot sequences staged onto the multishot sampler | | ext | extending an existing take | | music | music and lyric work | | image | stills (see README_IMAGE.md for the FLUX.2 writer) |

The brief

One craft instruct is always present. It is organised the way a shot is physically built, because that is the order the failures appear in:

WEIGHT    mass, gravity, contact, poses arrived at, joints in directions
TEXTURE   the physical particular over the impression; what shifts as it runs
LIGHT     direction, quality, one effect, one optical property
SOUND     everything audible comes from inside the shot
FRAME     camera placed once, distance, eye contact, cast, the closing image

Then a short law for each dial that is actually set, and nothing for the ones that are not. A bare shot reads as a clean brief rather than a stripped one.

The request always outranks a dial. Anything named in the intent wins over anything a dropdown seeded.

VOICE — register, grain, pace, and the scene setting the delivery — ships with any talking shot and always before the accent law.

Where the laws live

brain.py         the craft instruct, first person, speech, assembly, budgets
accents.py       49 accent notes and the accent law
wardrobe.py      appearance — viewer look, cast look, wardrobe, undress
closet.py        the garment catalogue and its pictures
dials.py         style, camera, transition, music, format
celebrities.py   face keys, rebuilt from h3_known_faces.json
worldscapes.py   settings and worlds
scenarios.py     scenario banks
artists.py       artist and vocal profiles
refs.py          the reference rail — what is attached, and what it is for
vision.py        one vision policy: sizes, per-mode caps, frame budgets

Plumbing: backend.py (LM Studio transport, streaming, abort), routes.py (panel endpoints), h3_studio.py (the node), imaging.py, negative.py, vram.py.

Dials

video_mode 8 modes · pov off / male / female · accent 49 + strength (natural / strong / thick) + a separate partner accent + who leads + spoken language (english / own) · dialogue 0–100% · wardrobe 591 garments in 16 categories + seed + undress · style 570 in 20 groups · style_look 176 · camera 40 · transition 11 · music 51 · scenario 107 · worldscape 57 · celebrity 503 · ref_recipe 7 · fps, seconds → beats and frame count · words_per_second short / medium / long · fmt 5 formats · lexicon names and terms.

Accents

49 accents, and the law splits them by what can actually be written down:

An accent whose note names a writeable shift (th toward d, dropped h, rolled r) gets SPELL IT — the accent goes in the letters of the words.
An accent whose note names only rhythm, stress or vowel colour gets grammar, word order, tags and address terms instead, and respelling is forbidden. Told to respell anyway, a model invents fake phonetics or borrows a louder accent's — which is how an Indian-tagged line came back French.

The accent is named inside the dialogue bracket, where it reaches the voice (<d>[English with heavy Ugandan accent] She is de queen.</d>), and nowhere in the prose attribution. Two speakers each get the full law, never a paraphrase.

Nodes

53 register in total — this pack's own plus the vendored packs below. They appear under LD / PromptMaster and as H3 … entries. The spine is:

🎬 H3 Studio → Unpack / Pack In → your H3 or LTX graph

with routing (H3 Route, H3 Route AV, Group Switch), samplers (H3 Multishot, H3 Chain, H3 Infinite Take), references (H3 Refs, H3 Ref Folder, H3 Auto Refs), audio (Music Studio, H3 Lock Audio, Audio Duration) and output (H3 Join, H3 Concat A/V, Video Grade).

Backend

LM Studio only. Point it at an OpenAI-compatible server — the default is http://127.0.0.1:1234 — and pick your model in the cog. The modes that look at pictures (ref, i2v, mv, multishot) need a vision-capable model loaded; set it as the vision model in the same panel.

Setting up LM Studio

Download and install LM Studio (Windows/Mac/Linux, free).
In the search tab, pull a model. This pack is built and tested against qwen3.8-27b-uncensored-hauhaucs-aggressive-mtp — search for it by name and grab the quant that fits your VRAM.
Before you load it, open the model's load settings and:
- Turn "Thinking" off (reasoning mode). A reasoning model spends its whole reply on a <think> block before writing anything usable, which wastes your token budget on scratch work instead of the shot script and can make the panel stall waiting for text that never shows up.
- Set Context Length to at least 32,768 (32k). LM Studio's own default (4k–8k depending on model) is too small for this pack's longer briefs — multishot, reference-heavy scenes, and accent + wardrobe stacked together can all run past it, and a context that's too short cuts the script off mid-sentence instead of failing cleanly.
Open the Local Server tab (left sidebar) and start it with that model loaded. Leave the port at the default 1234 unless you have a reason to change it.
In the H3 Studio panel's ⚙ settings: set server_url to http://127.0.0.1:1234, pick your model from the list, and set the panel's own ctx field to the same number you set in step 3 (32768). This ctx is a separate number the pack uses to estimate whether a prompt will fit — it does not read LM Studio's setting automatically, so the two have to be kept in sync by hand.

The managed llama-server backend was removed: it spawned llama-server.exe against a GGUF picked from disk, and it kept breaking for people who were connected to LM Studio anyway.

Connection settings live in cpld_conn.json, which is local and never committed.

Install

Copy this folder to ComfyUI/custom_nodes/PromptMasterLD/ and restart.

Assets are not in the repository. RefBank/ is excluded entirely and Wardrobe/ ships two example garments so the closet has something to resolve — the rest of both folders stays on the machine that made them. See .gitignore.

workflows/ ships an example graph — drop it into ComfyUI/user/default/workflows/ to load it from the workflow menu.

Tests

python selftest.py     # wiring, no LLM
python eval.py         # real shots through the live backend
python eval.py --dry   # build every brief, call no LLM
python eval.py ref     # only cases matching "ref"

selftest.py checks that dropdown keys map to real laws in both directions, that a set dial changes the brief and an unset one does not, that no two live laws contradict each other, that the arithmetic holds, and that every call routes.py makes into a law module binds against the real signature.

eval.py grades the generated script, which is the only layer that can see an accent bleeding into the wrong mouth, a second speaker coming back flat, two people's actions blending, or forty actions crammed into twenty seconds. Needs the backend up. The graders are blunt and mechanical on purpose — no model judging a model — and each failure hands you the offending line.

Licence

MIT — see LICENSE.

Bundled third-party packs

PromptMasterLD ships two packs inside vendor/ so the whole rail installs as one repository. Each keeps its own folder, its own __init__.py and its own LICENSE, unmodified, and is loaded by this pack's __init__.py.

| Pack | Author | Licence | Why it is here | |---|---|---|---| | ComfyUI-H3-Multishot | RiftCast / jlucasmcrell | MIT | upstream for h3ms_bridge; its instruct and sampler are what multishot is built on | | ComfyUI-H3-Motion-Context | NikoDemon80 | GPL | H3 Master's clip chaining |

Full credit to both authors. The multishot contract in dials.py is ported from jlucasmcrell's instruct and is marked as such in the source.

Licence note: ComfyUI-H3-Motion-Context is GPL-licensed. Bundling it means this distribution as a whole is subject to the GPL. If you would rather keep this pack permissive, delete vendor/ComfyUI-H3-Motion-Context/, remove it from _VENDORED in __init__.py, and install it separately — H3 Master already raises a clear message when it is absent.