Download for Windows · Mac (Apple Silicon) · Mac (Intel) · Linux (AppImage) · What you need · Release notes · Infinite Variance

IV Studio

A free Infinite Variance desktop app for making short films with
MiniMax H3 on your own PC. Give a recurring actor a face
and a voice, write a shot, hunt a few quick previews and finish the best one at full size. Make
pictures, songs and spoken lines, cut it all together and export. IV Studio drives your own
ComfyUI; nothing leaves your machine unless you choose the
cloud. No strong graphics card? Shots can also render on a GPU in your own RunPod or
Modal account, billed by the second, or on Comfy Cloud (see Cloud).

IV Studio 1.1.7 opens on a simpler welcome that asks how much you already know and keeps things short if
you're new, shows a picture for every one of the 165 Cinematography setups, lets Hermes write your prompts,
and opens Create's examples and staging board as pop-ups. 1.1.6 lets you name your agent (and tell it your
name), pick a colour for the whole app (red, blue, purple, orange or a dark grey), and gives the editor's cut
tool a red katana.

IV Studio 1.1.4 adds Hermes to the chat beside Claude Code and Codex, and a History tab for earlier
conversations. The IV Studio plugin for Hermes works the other way
round: Hermes's own video_generate and image_generate render in IV Studio, on your GPU, for free.

IV Studio 1.1.3 lets you talk to the chat (it hears you with IV Studio's own speech recognition, on your PC) and
read its replies aloud, shows every agent's work in an Activity log with Undo, gives the chat a cleaner look and
its own window, and tells it more about what you're looking at.

IV Studio 1.1 brings your own Claude Code or Codex into the app as a chat that can make, edit and fix things
for you, MiniMax H3 Image and Seedance 2.5 beside MiniMax H3, effects for any clip, nearly every feature on your
own RunPod or Modal, speed presets for smaller graphics cards, and a new look that fits any window. See
what's new in 1.1.

IV Studio is Powered by MiniMax H3. Using H3 through it means agreeing to the
IV Studio terms of use, which carry MiniMax's use restrictions: IV Studio asks once, before your first
H3 render. To report misuse, email [email protected].

Create: Tess's rainy bus-stop shot with its H3 prompt, the checker's note, the prompt tools and the pinned Your shot card

What's new in 1.1

Your own Claude Code, Codex or Hermes, inside IV Studio

Chat. The Chat button in the title bar opens a conversation with your own Claude Code, Codex or Hermes, on
your own plan, with every IV Studio tool: it can look things up, write and render shots, sequence, edit, make
music and voices, and answer questions about your projects. It knows which page you're on and which shot,
film or edit is open. Anything that deletes, costs money, installs or changes a setting shows an Allow /
Don't allow card first. Conversations are kept: the History tab (1.1.4) lists them, newest first, with which
agent and when, and opening one carries on where you left off. On a wide window the
chat sits beside the page; on a smaller one it slides over it. Works on Windows, Mac and Linux.
Hermes (1.1.4). Pick Hermes beside Claude Code and Codex and your own
Hermes Agent works in the chat, on the model you chose in Hermes. IV Studio
talks to it through Hermes's Agent Client Protocol server (hermes-acp, found where Hermes's installer puts it),
gives the session IV Studio's tools, and resumes the same Hermes session for each message. Anything Hermes asks
approval to run outside IV Studio (a shell command, a file edit) is declined and shown as a declined step.
Talk to it (1.1.3). Press the mic in the chat, say what you want and press send. IV Studio's own Whisper writes
it down on your PC (Audio → Voices → Set up voices once). Talking while the agent works stops its answer and
sends your new words; renders it already started keep going. The speaker button reads each reply aloud in one of
your voices (the first one takes a while as the voice loads; after that a few seconds).
Its own window (1.1.3). The arrow in the chat's header opens it in a window of its own: keep it on top while you
work, and put it back in the app when you're done. It still knows what's open in the main window, and its links
open things there.
What it can see (1.1.3). Besides the page, each message tells the agent what you picked on it (a shot's version,
the clip and playhead in the editor, the item open in the Library, how Create is set up), what's rendering or
waiting, and the last render that failed in the past half hour.
Activity (1.1.3). The chat's Activity tab lists everything agents did in IV Studio: the chat, Dashboard ideas,
and Claude Desktop, Codex, Claude Code or Hermes using IV Studio's MCP server from outside, each by name. Steps that
change something, spend or fail are kept for 30 days (looking things up isn't), with a link to what they made and
Undo this change for any agent's change to an edit. Agents can read it too (list_activity).
Working memory (1.1.3). The chat's header shows how full the agent's context window is (exact for Claude Code;
estimated for Codex from its usage) and, when clicked, the chat's tokens. A note shows when the agent summarizes
earlier messages to make room. Your plan's 5-hour and weekly limits aren't shared with apps, so they aren't shown.
Ask about this. Buttons around the app hand the chat a question: Ask … to fix it on a failed render
(with IV Studio's own explanation), Ask … on any run and Ask … to pick the best take on the Shot page,
and Ask … about this in the Library viewer.
Drop files in. Drop pictures, videos, sound or a script onto the chat, or use its paperclip. The agent can
look at a picture or a moment of a video, and use the files as a first frame, a reference, a song or a script.
Show me where. Ask "where do I train a character?" and IV Studio goes to the right page and lights up the
right button, with the agent's words beside it.
Edit by talking. On an open edit, press Ask … and say "tighten the middle", "add a song under it" or
"put captions on". Captions are new: Whisper hears each spoken clip and the words go on as short lines at the
bottom, timed across the clip (silent clips are left alone). Every change the agent makes gets an Undo
this change in the chat, even when the edit isn't open, and an open edit shows the change at once.
Recipes. When something works, say "save this as a recipe". It becomes a button on the Dashboard that asks
the chat for the same thing again, written so it works another day.
Overnight batches. "Make 10 variations of this idea overnight, three takes each": the agent writes each
variation and queues them on your PC (free, up to 12). IV Studio keeps the PC from sleeping, ranks the takes
as they finish, and leaves a report in the chat with the best ones to open.
A short film in one conversation. Make one with … on Projects: the agent takes an idea through script,
cast, storyboard, shots and rough cut, and stops for your OK at each step. It only starts filming after you
approve the storyboard.
A spending limit. Settings → Agents: let the chat spend Comfy.org credits without asking, up to a daily
amount and an amount per step (off unless you turn it on). Each paid step shows what it spent and how much of
today's limit is used. Renders on your cloud GPU, training and anything without a known price always ask.
Seedance now works from the chat too.
Let Claude Code make it. On the Dashboard, a button beside Make it hands your idea to your own agent:
it writes the prompt and renders the previews on this PC, with its steps in a panel and a Stop button. It
never deletes, installs or spends.
Install them from IV Studio. The welcome asks which agent you use (Claude Code, Codex or Hermes) and what to
call it, and Settings → Agents can install Claude Code or Codex with its maker's own installer and open its sign-in.
Your agent's name (1.1.6). Name your agent and tell it yours, in the welcome or Settings → You and your agent.
It answers to that name in the chat, on the Dashboard ("Let Nova make it") and in the activity log.

The chat beside the Dashboard: Claude Code answering which shots aren't finished yet, film by film, after looking them up with IV Studio's tools

Make more

MiniMax H3 Image. A new picture model in Create → Image: big, detailed pictures at 4, 8 or 16 MP that
follow long descriptions closely, with MiniMax's own prompt tips beside the box. It needs ComfyUI 0.45 or
newer and a 22 GB download (Settings → Setup). New pictures are clean; edits with the int8
download come out over-sharpened for now.
Seedance 2.5 in Create. A model switch beside MiniMax H3: the same shots on ByteDance's Seedance 2.5,
through ComfyUI's partner nodes, on your Comfy.org credits. Cheap 480p drafts to try ideas, or 720p and
1080p takes, 4 to 30 seconds. The price shows on the button and you confirm before anything is charged.
Write it for me writes for Seedance from a Seedance 2.5 prompt guide.
Your Comfy account. Settings → Comfy account holds your Comfy.org key and shows your credits, also in
the title bar. Anything paid is priced first, from ComfyUI's own prices, and needs your OK.
Free finishes for Seedance takes. Finish a draft on your own graphics card: SeedVR2 (cleanest), DLSS 5
(quickest), or H3 redrawing it at full size followed by DLSS 5 or SeedVR2.
Upscaler strength. DLSS 5 has a strength and a look (Default, Natural, Cinematic), SeedVR2 an extra-detail
slider and H3's redraw a strength, in the Finish dialog and in Enhance.
Effects on any clip. Right-click a clip in the editor → Effects…: Dub (new words in any language, with new
lip movement), Clean plate, Day to night, Add water, SDR to HDR, Remove background and Smooth or slow motion.
The result takes the clip's place, and Undo brings it back. Built on
Freefield's apps (MIT).
Effects, Sounds and picture tools. Workflows → Effects turns a photo into an effect clip, a transformation
or a transition. Audio → Sounds makes sound effects, ambiences, one-shots and loops with Stable Audio 3.
Create → Image → Tools inpaints (paint over a part and describe it), relights, removes backgrounds, upscales
and makes depth or normal maps.
LoRAs for every model. The LoRAs page and every LoRA picker cover MiniMax H3, LTX-2.5, Wan 2.1, Flux 3
Klein, Flux.2 Klein, Krea 2, Qwen Image 2.1, Z-Image Turbo, SDXL (with Pony and Illustrious) and Moss: find a
model's LoRAs on Civitai, see yours sorted by model, and send one to that model's tray. A LoRA IV Studio
can't place can be assigned by hand.
Start from an example. Create's Start from an example button opens six ready-made shots in a pop-up;
each fills in a working prompt. The welcome wizard's Make your first shot opens the first one ready to hunt.
Speed presets. Quality, Balanced, Fast and Low memory trade picture size and steps for time and memory.
Check my PC picks the one that suits your card; every preset fits a 12 GB card, the 2 MP finish too.
How long, and what it costs. Hunt and Finish say how long they'll take on your machine, worked out from
your own earlier renders, and what a cloud render will cost.
Failed renders in plain words. Out of memory, a missing model or node pack, a full drive, credits and more
are explained with a button for the fix. ComfyUI's own message is still under Details.
Compare and finish. Mark up to four takes, finals or enhanced copies and play them in sync, side by
side, as a wipe or in a grid, zoomed in on the same spot. Picking a take opens the Finish settings over the
page, and every finished version sits in one card.
Library details. Open any render to see how it was made: model, size, length, seed, character, LoRAs,
how it was finished and enhanced, render time, cost and the prompt (with Copy). Tiles show size, length and
model; search finds words in prompts.
An end card, if you want one. The editor can add a "Made with IV Studio" end card to an export (Edit
settings → End card), off unless you turn it on, and each export shape says where it fits.


_{Seedance 2.5 beside MiniMax H3, priced before you hunt}	_{Right-click a clip for Effects}

_{Compare up to four versions in sync}	_{Every render says how it was made}

Catch problems sooner, finish anywhere

Takes that cut away. H3 sometimes cuts to a close-up nobody asked for in the middle of a one-shot take. The
take card now says "cuts away at 2.4 s", the take scores lower, and Use 0.0 to 2.4 s keeps the part
before (or after) the cut in an edit; Carry on from it starts a Continue shot from that part, cut to H3's
frame grid. Only hard cuts are caught, not slow dissolves.
First and last frames held. In First / last frame mode each take says whether it really opens on your
first picture and ends on your last.
"No …" wording. H3 has no negative prompt, so "No photographs appear anywhere" can hang photographs on the
wall. The checker lists every "no / without / never / not …" phrase with what to write instead. IV Studio's own
tested lines ("No one speaks.", "There is no writing anywhere in the picture.") stay.
Premiere / Resolve. The editor's Premiere / Resolve button: the timeline goes out as FCP 7 XML (Premiere Pro and DaVinci
Resolve), FCPXML 1.10 (Resolve, Final Cut Pro) and a CMX3600 EDL of V1, with every clip copied into a Media
folder beside them and the folder zipped to send. Tracks, trims, pictures, volume, hidden, muted and locked
tracks and markers come over; titles come over as markers (and as text in the FCPXML). Fades and title
styling don't.
Recommended LoRAs. LoRAs → Recommended (MiniMax H3): Camera Motion, Natural Face Speech, Motion
Continuity Repair, Spatial Physics, Head-mounted FPV, 360° Orbit and Transformation, each fixed to one tested
file and checked when it downloads, with what it does, the modes it works in, how to prompt it and its
licence. Use in a shot puts it in Create's tray with its trigger and strength.
Turnaround sheets. A character's settings make five views of them turning around from their face
picture, with matlod's Turnaround LoRA and the H3 Contact Sheet nodes (both downloads you choose).
Search by what's in a file. While nothing renders, your local prompt writer (Ollama, or your
OpenAI-compatible server when it sees pictures) looks at each new video and picture in the Library and writes
a sentence and a few tags, so search finds "red car" or "beach at night". It stops as soon as a render starts;
Settings → Prompt writer turns it off.
Soft mouth. A reference video in References mode (and Replace's Swap people) can have its mouths blurred
first, so the new face doesn't mouth the words spoken in it. Off while the shot has a spoken line. It uses the
face models from Characters. H3 can copy the blur into a take's first second; trim it in Edit.
Dailies. When an overnight batch finishes, Videos\IV Studio\Dailies gets a reel of every take, the best
of each shot first, each after a card with its seed and checks, and notes.md on what to look at first. The
chat's report has Watch the dailies and Read the notes.

These ideas come from keys and Sam Wasserman's Imagination Studio (Apache-2.0); the code is IV Studio's own.

Your own cloud does nearly everything

Almost every feature on your RunPod or Modal. Pictures and picture tools, storyboard sketches, character
design, Music, Sounds, Cut out, Reframe, Replace (with Remake), Effects, Retake in the editor, the SeedVR2
finish and Enhance now run on your own cloud GPU, not just shots. The models a render needs download into
your cloud storage the first time, after you agree to the size; your LoRAs and RefMods go up once.
Voices in your cloud. Voices, Lines, Speak it and Whisper's read-backs run on your RunPod or Modal too.
Hugging Face models. A Hugging Face token on the Cloud page (kept sealed) lets the cloud fetch gated models
like LTX-2.5 and Flux.2 Klein.
Cloud setup you can follow. Setting up RunPod or Modal shows each step with a timer and the time left,
and tells you when it's ready (1.0.6). Modal's sign-in, sign-out and new accounts work (1.0.3 to 1.0.5).
The IV Studio cloud beta. Want IV Studio to run the cloud for you? Join the waitlist on the Cloud page or
in Settings → Privacy. Only the email you type is sent, and only when you press Join.

Easier to use

A simpler welcome (1.1.7). The first thing IV Studio asks is how much you know about making AI video and
pictures. Then Make it yours (your name, your agent, a colour, easier-to-read options), a look at your PC,
where things render, what to download and you're ready: six steps on one row. If you're new to it, IV Studio
keeps each step to a couple of plain sentences and folds the details away with its suggestions already set.
Your colour (1.1.6). Red is IV Studio's own; blue, purple, orange or Dark (greys and white) recolour every
button, highlight and the artwork, in the welcome or Settings → Accessibility.
Cinematography you can see (1.1.7). The Cinematography button (with its little cinema camera) opens all
165 setups as pictures of one scene, a woman in a red coat at a rainy café window, shot each way: film stocks,
lenses, lighting, camera moves, moods. Hover one to read the phrase it adds; camera moves are marked.
The staging board in a pop-up (1.1.7). Stage it on a board on Create opens the board large; Use this
staging puts a one-line summary under the prompt, with Edit to open it again.
Editor tools you can see (1.1.6). The Razor (C) turns the cursor into a red katana with a guide line, and
each cut flashes a slash (not with Reduce motion). Select, Ripple and Roll have their own cursors and icons.
A new look. Tall animated banners with a quick-start bar on every main page, and an Audio front page.
Nothing flashes (every banner is checked against the WCAG flash limit), animations pause when they're off
screen, and Reduce motion holds them still.
Fits any window. Pages, banners and the title bar adapt to smaller windows, and choice cards wrap
instead of being cut off.
Searches that say so. The search bars in the banners have a Search button that takes you to the
results and a count of what matches; + New shot, + New film and + New edit are their own buttons.
What's new as a pop-up. After an update, a small See what's new in this update button opens this list.
Simple words. Check my PC asks how much you know about AI video. New to it? IV Studio says
"trained look" for LoRA, "variation number" for seed, "quick sharpen" for DLSS and so on. Only the words
change; switch it in Settings → Words.
Get ComfyUI for me. No ComfyUI yet? Setup downloads ComfyUI's official portable build for your graphics
card into a folder you choose, checks it, unpacks it and starts it. Already have one? IV Studio keeps using it.
A tidier Setup. A status card first, then folding sections that open by themselves when something needs
you. Restart ComfyUI is one click (it asks first while renders run), and node packs that are installed
but not loaded yet say so.
Less waiting on memory. Effects free the GPU before their video model loads, and reopening IV Studio while
ComfyUI starts no longer starts a second one.
An installer in IV Studio's colours, with the eye down the side.
Scroll bars in the theme, and no clipped corners (1.1.6). Every page was checked from 1100 × 720 to
2560 × 1440 with the chat open and closed.


_{Name your agent, tell it yours and pick a colour}	_{Every cinematography setup as a picture}

Updates, privacy and credits

Updates on a Mac. Mac copies now download and install new versions themselves, like Windows and Linux:
Restart to update, or it installs when you quit. (IV Studio needs to be in your Applications folder.)
Privacy. Settings → Privacy: IV Studio sends nothing about how you use it unless you turn on Help improve
IV Studio, which counts only opens and renders, never prompts, pictures or files. Incognito never counts.
Training on Fizgig 9. Character training uses Fizgig 9.3.2, on this PC, Modal and RunPod. RunPod training
pods start from Fizgig's own RunPod template, so your training time supports Fizgig too.
Credits. Settings → About → Credits lists every model, tool, node pack and library IV Studio is built on:
who made it, what it does here and its licence.
Terms for MiniMax H3. Before your first H3 render or training, IV Studio shows the rules for H3 in plain words
and asks you to agree to its terms of use once (they carry MiniMax's licence restrictions and Acceptable
Use Policy). Read them again, or report misuse, in Settings → About. If you're updating from 1.0, you'll be asked
at your next H3 render.

What's new in 1.0

Help on every page. A How it works button on every page lights up its real controls one at a time
with a line on each. The welcome wizard offers a full tour of the app, and Settings → About takes it again.
Check my PC. On first launch IV Studio reads your graphics card, memory, processor and drives (nothing
leaves your PC), says plainly what it can do, and sets previews, finals, the prompt writer and downloads to
suit. A guided cloud setup connects RunPod, Modal or Comfy Cloud in a few steps.
Redesigned. Create (video and image), Settings, Characters, Voices, Lines and Music are numbered step
cards. The Workflows tools (Sizzle, Reel, Sequences, Shinobi and Replace) open on animated stages with their
main bar centered.
Find anything. Shots, Projects and the Library group your work by day and filter it by when, kind of
shot, who's in it and shape, in a panel that folds away. Long histories open at once: tiles load as you
scroll.
Train characters anywhere. Free on your own graphics card (Fizgig), or on your RunPod or Modal account,
with presets, a spending cap, Train more and likeness scoring. Or make a RefMod character from a few
pictures in seconds.
More control. Pick a shot's sampler and scheduler, and keep a resizable live preview of any render
floating over every page.

What it does

Dashboard. IV Studio opens on one-click starts for everything it makes, your latest renders and
your cast. Every page has its own banner art and its own How it works tour.
Projects. A short film as a six-step wizard for anyone: an idea, a script (written for you or your
own), the cast and places, a storyboard of drawn sketches you can play as an animatic, filming every
panel, and the finished film. Faces and rooms stay the same from shot to shot.
Shots. Write a shot (or type the idea in plain words and let Write it for me write the H3
prompt; it looks at your reference pictures and videos first), hunt one to three quick previews, pick
one and finish it at 1 to 2 MP with DLSS, an RTX 2× upscale and frame generation, or with SeedVR2. Four
modes: text, reference pictures, video and audio, first and last frame, and continuing a video. Create
lays it out as numbered steps beside a pinned Your shot card.
Directing. A prompt checker that catches H3 mistakes before you spend a render, @ mentions for your
cast and places, live step pictures while takes render (also in a floating preview over every page),
takes ranked by words heard, likeness and sharpness, side-by-side Compare, a Review that names each
problem with its seconds and one fix, and a top-down staging board that becomes exact screen directions
or a moving grey previs.
Images. Create switches between Video and Image: make or edit pictures with eight image
model families (Flux.2 Klein, Flux 3 Klein, Krea 2, Moss, Qwen Image 2.1, Z-Image Turbo, SDXL and H3
stills with your H3 characters), then use one as a shot's first frame.
LoRAs. Browse Civitai's MiniMax H3 LoRAs as tall picture cards whose clips play on hover, open one
for a full-size gallery of its examples, and download it straight into ComfyUI. Your own LoRAs show as
pictures too, and Create's LoRA tray adds them to a shot with a strength slider.
Characters. Recurring actors as trained H3 LoRAs with a trigger word, or as RefMods made from a few
pictures in seconds. Use the ready-made Tess and Theo, add a LoRA or RefMod you have, or design, clip and
train your own: free on your graphics card, or on your RunPod or Modal account.
Voices. Give each character a voice: design one from a description, or save one from a recording.
Type a line and Speak it: the character says it on screen, lip-synced. Runs on your PC with
VoxCPM2 and Whisper.
Music. Songs with sung lyrics, or instrumentals, from a style and lyrics: MiniMax Music 3 for full
songs up to five minutes, ACE-Step 1.5 for a draft in seconds. Put a song under an edit, or have a
character sing it.
Edit. A Premiere-style editor with a full-screen workspace: project bin, source and program
monitors, as many video and audio tracks as you need, filmstrips and waveforms, markers, J/K/L,
insert and overwrite, razor, ripple and rolling trims, and MP4 export at 1080p, 4K, vertical or square
at a set loudness. Right-click to make new footage on the timeline (extend, bridge, fill a gap, retake a
stretch), cut out a person or thing with see-through edges, or reframe a shot to a new shape with H3
drawing what lies beyond its edges.
Workflows. Five tools, each opening on its own animated stage. Sequences: several shots as one
continuous video. Sizzle turns a website into a promo film: your agent storyboards it from the real
site, then H3 shots, pictures, music and the site's own screens are cut together on the beat. Reel
cuts pictures or clips to a song's beat. Replace swaps a person or changes the scene in a video while
the camera, motion and sound stay. Shinobi lets an agent cut a whole edit in code.
Library and Queue. Every shot, sequence, picture, export, song and spoken line in one place, a day at a
time, filtered by when, where it came from, kind, who and shape. Above it, the render queue shows what's
rendering and what waits.
Help. How it works on every page, a full tour from the welcome wizard, What's new after every
update, and the PDF guide.
Check my PC. Reads your hardware on first launch and sets IV Studio up to suit it.
Cloud. Render shots and sequences on your own RunPod or Modal account, or on Comfy Cloud, set up with a
guided wizard, with a running time and cost meter in the title bar and a Shut down GPU button.
Incognito. A private space whose work never shows in the normal app.
Agents. A local MCP server with 214 tools and a start menu gives Claude Code or any MCP client nearly
everything a person can do in the window.
Updates. IV Studio updates itself: one click installs a new version and reopens.


_{Start anything from the Dashboard}	_{How it works: every page walks you through itself}

_{Check my PC sets IV Studio up for your computer}	_{Hunt a few takes, review them, pick one and finish it}

_{Every shot, a day at a time, with filters}	_{A short film, step by step}

_{Pictures too: generate, edit or fix on nine model families}	_{Stage the shot on a board; it becomes exact directions or a previs clip}

_{Recurring characters, ready-made or your own}	_{Train free on your PC, or rent a cloud GPU}

_{LoRAs for every model, found on Civitai}	_{Every LoRA's examples, trigger words and licence}

_{A voice for every character}	_{Songs and scores, made on your PC}

_{A Premiere-style editor, with your agent cutting alongside}	_{Everything you've made, in one place}

_{Seven tools for after a render}	_{Sizzle: a website, made into a promo film}

_{Reel: pictures and a song, cut to the beat}	_{No big GPU? Render on your own cloud account}

_{Settings, one card per topic, with speed presets}	_{Setup checks and downloads what ComfyUI needs}

What you need

Windows 10 or 11 (64-bit), 64-bit Linux (the AppImage, from 0.2.8; see Linux), or a Mac
with macOS 12 or newer, Apple Silicon or Intel (from 0.3.2; it renders in the cloud, see Mac).
An NVIDIA GPU, or the cloud. 24 GB of VRAM is what IV Studio is tested on (an RTX 4090); 16 GB works
with smaller previews and finals. 64 GB of system RAM is recommended. Without a suitable card, nearly
everything can render on a cloud GPU in your own RunPod or Modal account, voices included (see
Cloud).
ComfyUI, version 0.33 or newer, running on the same PC.
IV Studio talks to it at http://127.0.0.1:8188 (change it in Settings → ComfyUI; localhost:8000 for the
ComfyUI Desktop app is fine, and Check my PC finds a ComfyUI on another usual port by itself). It can start ComfyUI for you: it
finds your ComfyUI shortcut, the ComfyUI Desktop app or a portable install's run_nvidia_gpu.bat (or you pick one
in Settings → ComfyUI, or pick your ComfyUI folder and the start file in it), starts it when IV Studio opens, and
offers Start ComfyUI whenever it isn't running. No ComfyUI yet? On Windows with an NVIDIA card,
Settings → Setup → Get ComfyUI for me downloads ComfyUI's official portable build (about 2 GB, its own Python
inside, nothing else installed) into a folder you choose, checks it against the checksum ComfyUI publishes,
unpacks it and starts it. It picks the build for your card (the CUDA 12.6 one for 10-series and older) and says
if your driver needs updating first.
Disk space. About 60 GB for the H3 models. Music adds about 14 GB (MiniMax Music 3) or 10 GB
(ACE-Step); voices add about 12 GB. Each is a one-click download inside the app.
Optional: Ollama for the free local prompt writer;
Fizgig to train characters on your own graphics card (16 GB or more);
and a RunPod or Modal account for cloud rendering and training.

Install

Download IV-Studio-Setup.exe
(about 180 MB).
Run it. The installer isn't code-signed yet, so Windows may show "Windows protected your PC".
Click More info → Run anyway. It installs for your user only; no admin rights needed.
Open IV Studio. On first launch Check my PC reads your graphics card, memory and drives (nothing
leaves your PC), says what it can do, and sets previews, finals, the prompt writer and the downloads to
suit; it ends by offering a two-minute tour of the app. IV Studio starts ComfyUI for you when it can.
Before your first MiniMax H3 render it asks you to read and agree to the terms of use once.
Settings → Setup checks ComfyUI, your GPU, the node packs and every
model, installs missing node packs into ComfyUI with one click (restart ComfyUI afterwards), and
downloads missing models straight into ComfyUI's models folder (each checked against the SHA-256 its
host publishes). Its Updates row keeps ComfyUI, its node packs and ComfyUI-Manager up to date.
For voices, open Audio → Voices and press Set up voices once. For music, open
Audio → Music and download the engine you want.

Your projects, voices and songs are kept in %APPDATA%\IV Studio; renders are saved in ComfyUI's
output folder under H3\.

Linux

Download IV-Studio.AppImage
(about 240 MB), make it executable (chmod +x IV-Studio.AppImage) and run it. AppImages need FUSE
(libfuse2, called libfuse2t64 on Ubuntu 24.04).
IV Studio finds a ComfyUI folder in your home folder (or pick it in Settings → ComfyUI) and starts it with
a .sh script beside it, or with its main.py and the folder's own venv Python. ComfyUI's output goes to
comfyui.log in IV Studio's data folder, ~/.config/IV Studio.
Saving API keys (Civitai, RunPod, Comfy.org) needs a keyring (GNOME Keyring or KWallet).
Shinobi, previs and website capture use Google Chrome, Chromium or Microsoft Edge.

What's different: DLSS 5 is Windows-only, so finals use a 2× AI upscale (Real-ESRGAN) and RIFE frame
interpolation when those are installed in your ComfyUI, as in the cloud. The Linux build is built and checked on
GitHub's Ubuntu machines (every page, the agent server, Shinobi and website capture); renders on a Linux GPU
haven't been tried yet. In Settings → Agents, the command for Claude Code or Codex runs a small script that opens
the AppImage, so it keeps working after updates.

Mac

A Mac has no NVIDIA graphics card, and MiniMax H3 needs one, so on a Mac IV Studio is the studio and the cloud is
the GPU: shots, hunts, finishes and sequences render on your own RunPod or Modal
account, or on Comfy Cloud.

Download the .dmg for your Mac: Apple Silicon (M1 and newer)
or Intel. Open it and drag
IV Studio into Applications.
The app isn't signed with an Apple Developer ID yet, so the first time macOS says it can't check it. Open
System Settings → Privacy & Security, scroll down and click Open Anyway next to IV Studio (once).
Open the Cloud page (the Dashboard and the title bar point there) and set up RunPod, Modal or Comfy Cloud.

What's different on a Mac:

Renders: in the cloud. Finals there use a 2× AI upscale (Real-ESRGAN) and RIFE frame interpolation in place
of DLSS 5, which is Windows-only.
Pictures and music need a ComfyUI with their models: Comfy Cloud, or a ComfyUI on your Mac (IV Studio
finds and starts the ComfyUI Desktop app, ComfyUI.app, or a ComfyUI folder in your home folder). Note that
the ComfyUI Desktop app listens on port 8000: set http://127.0.0.1:8000 in Settings if you use it.
Voices run on your Mac's processor, not a GPU, so they're slow; not yet tried on a real Mac.
Updates: IV Studio installs its own updates on a Mac too. macOS's built-in updater needs a paid Apple
signature, so IV Studio downloads the new version itself, checks it against the release's checksum, swaps it in
when you click Restart to update (or quit) and opens again. That needs IV Studio in your Applications folder:
a copy opened straight from the .dmg or from Downloads says so and opens the download page instead.
Copies before 1.1 still open the download page once.
Keys (RunPod, Civitai, Comfy.org) are kept in the macOS Keychain.
Your data is kept in ~/Library/Application Support/IV Studio; exported edits go to Movies/IV Studio.
ffmpeg in the Mac app is Martin Riedl's GPL static build of FFmpeg 9.0.2 (the npm package's Mac builds are
made with --enable-nonfree and can't be shared).

The Mac app is built and checked on GitHub's Apple Silicon and Intel machines (every page, the agent server,
edit export, Shinobi and website capture with Chrome). It hasn't been tried on a person's Mac yet.

Updates. From 0.2.1 on, IV Studio checks for a new version when it starts and every six hours, and
downloads it in the background. Click ↻ Update to … at the bottom of the sidebar (or Restart to
update in Settings → Updates) to install it and reopen; otherwise it installs the next time you quit.
It waits while an edit is exporting or models are downloading. Coming from 0.2.0 or older, install 0.2.1
by hand once.
The portable Windows copy updates itself the same way from 1.1.2 on: it downloads the new portable zip into
its data folder, checks it, and swaps the program files in place, keeping the data folder. A portable copy on
1.1.1 or older needs the 1.1.2 zip unzipped over it once.

Settings → Setup: a ready status card with Restart ComfyUI, and folding sections for models, node packs and optional tools

Dashboard

IV Studio opens on the Dashboard: one-click starts for a shot, a picture, a short film, an edit, a song,
a voice line, a website film or a new character; your most recent renders; and your cast, each with their
face, their voice, and New shot to start a shot with them already picked. After an update, a strip at the
top says what's new in that version (Settings → Updates keeps the last few).

Every page opens on its own banner, drawn in the Infinite Variance eye style: a clapperboard for Shots,
scissors for Edit, a storm cloud for Cloud, a record player for Music, and so on. On wide windows the art
stays against the right edge.

How a shot works

Create. Pick a mode, a character and write the shot.
- Text: just the prompt.
- References: pictures, videos or audio guide the shot (<Picture 1>, <Video 1>, <Audio 1>).
  Click a reference to crop or trim it; videos snap to H3's lengths. Size: Fit for H3 resizes a
  picture or video to H3's own sizes (about 0.5, 1 or 2 MP, both sides on a 32-pixel grid, video at
  24 fps), which some modes, like Retake, need.
- First / last frame: the clip opens on a picture and can end on a second one.
- Continue a video: carries a clip on from its last frames. Hunt 1 to 3 takes like any shot; the
  finish refines the picked take at full size and joins the source back on, so the final is the whole
  continued video.
Create is laid out as numbered steps (what kind of shot, who's in it, what it builds on, what happens,
shape and size) beside a pinned Your shot card that shows the frame in its shape and the Hunt button.
At the top, pick the model: MiniMax H3 on this PC (free) or Seedance 2.5 on your Comfy.org credits
(see Seedance 2.5). For H3, a Speed preset sets the preview size, steps, takes and the
finish together: Quality, Balanced, Fast or Low memory (Check my PC picks one for your card).
Set the aspect, length, preview size and how many previews to hunt (1 to 3). Optional extras: a voice
line to lip-sync, RefMods and LoRAs (both in Who's in it), and under Advanced the steps, sampler and
scheduler (from your ComfyUI's own lists; er_sde and beta by default).
Hunt. IV Studio queues the previews on ComfyUI and shows them as they finish.
Pick and finish. Choose a preview and set the final size (1, 1.5 or 2 MP), the enhancer and the
sound source. The finish starts from the preview's own latent, so the motion stays the same.
- DLSS 5 (the default): RTX 2× upscale and frame generation (30/50/60/120 fps), seconds per shot.
- SeedVR2 (3B): restores face, eye and hair detail and upscales to 1080p, 1440p or 2160p, on any
  NVIDIA card. Enhance runs either one again on a finished shot.
Picking a take opens the Finish settings over the page. DLSS 5 has a strength (its noise reduction:
higher adds contrast and skin texture, lower stays closest to the take) and a look (Default, Natural,
Cinematic); SeedVR2 has extra detail (0 restores faithfully; more redraws more, smoothing skin and
changing faces a little). Every final and enhanced copy sits in one
Finished versions card. Compare plays up to four takes, finals or copies in sync: side by side, as a
wipe or in a grid, zoomed in on the same spot, with frame-by-frame stepping.
From any finished version, Continue from this opens a new Continue shot with it as the source (same
character, LoRAs and shape), and Use the last frame saves its final frame as a picture in the Library,
to start a new shot from it as the first frame or to change it in Create → Image first.
Shots. Every shot keeps its runs, previews and finals. Show in folder opens the file. The Shots page
groups them by the day you last worked on them (older ones by month; each day's heading stays at the top as
you scroll and folds away with a click) and filters them by status, when, kind of shot, who's in it and
shape, in a panel that folds away. Tiles load their video only as they come near the screen, so a long
history opens at once.

Seedance 2.5

The model switch in Create also runs ByteDance's Seedance 2.5 through ComfyUI's partner nodes, billed to
your Comfy.org credits (add your key in Settings → Comfy account; your balance shows there and in the title
bar). It's the same Create page: text, references, first/last frame or continuing a clip, 4 to 30 seconds.

Quality. Draft (480p) takes are cheap: try a few, then finish the one you like. Or make 720p or 1080p
takes straight away.
The price first. The Hunt button shows ComfyUI's own price for the takes, and a confirmation asks before
anything is charged. Paid finishes ask too.
Writing. Write it for me follows a Seedance 2.5 prompt guide: references are named @image1,
@video1, @audio1 and each gets a job; a character comes along as their face picture.
Finishing a take. Seedance 1080p (paid), or free on this PC: SeedVR2 (cleanest, about 4 minutes for
5 seconds), DLSS 5 (about 15 seconds), H3 + DLSS 5 or H3 + SeedVR2 (H3 redraws it at full size first; faces
change a little), or keep the take as it is.

Before and after a render

Prompt checker. Before a hunt, IV Studio checks the prompt the way H3 reads it: the real length on
H3's frame grid, blank fields, the opening line a mode needs, cut times, too many words for the seconds,
quotes and speaker names around spoken lines, references named but not attached, a LoRA character's face
described in words, filler words and unknown @ names. Most findings have a Fix; errors hold the hunt
until you fix them or choose Render anyway.
@ mentions. Type @ for your cast and places. The picked character becomes {char}; anyone else with a
picture is attached as a reference, and a text shot switches to References mode by itself.
Draft or Standard. Draft previews take 4 steps instead of 8, roughly 40% faster, for hunting ideas.
Live previews. Each take's step pictures stream to the page while it renders (H3's tiny decoder
through KJNodes), so a bad take shows itself in about 20 seconds. Live preview (above the Library
button) floats the running render over every page: drag it to any corner, resize it, or fold it to a pill.
Ranked takes. When a run finishes, the takes are scored: the spoken words Whisper hears back, likeness
to the character (InsightFace), no frozen stretches and sharpness. The best is marked only when it clearly
leads, with the reasons in words, and every pick you make tunes the weights.
Runs as versions, and Compare. Each run shows what changed since the last, the prompt as written and as
rendered, and a word diff. Compare plays two takes side by side or as a wipe, in step. Edit and run
again renders into the same shot.
Review a take. Your prompt writer watches six frames of a take (and reads what Whisper heard), names
each problem with its seconds, says what works and picks one fix: revise the prompt, change a reference,
split the action, use a clean stretch, or fix it in the edit. It remembers the shot's earlier reviews, so
a problem that came back after a rewrite gets a different kind of fix.
Staging board. A top-down floor with up to four people (where they face, who they look at, where they
walk) and the camera (lens, height, ten moves). It becomes exact screen-side directions in the prompt, and
warns when a cut crosses the 180-degree line. For arcing moves, Make a previs clip renders the board in
motion as a grey clip that H3 follows as <Video 1>.

The directing rules the writers and the review follow (purpose first, staging by screen side, no padding,
frame the action) are Matthew Hollis's own method, built into src/service/directing.js.

Write it for me

You don't need to know H3's prompt format. Type the idea in plain words, for example "Tess in a parked
car at night admits she kept the letter", and press Write my prompt. You get a finished H3 prompt
back to edit, keep or undo. It always uses your mode, length, aspect, character and media. Choose the
writer under Settings → Prompt writer:

Local (free): a model in Ollama on your PC with IV Studio's H3 rule sheet. The recommended model is
huihui_ai/qwen3-vl-abliterated:8b-instruct (about 6 GB, one click to install), huihui-ai's uncensored
copy of Qwen3-VL 8B Instruct (Apache-2.0), so it doesn't refuse mature ideas. Any other Ollama model works
too, including the original qwen3-vl:8b-instruct. While ComfyUI is rendering it runs on the CPU, so it
never takes H3's VRAM. Ollama somewhere else (another container or PC) works too: set Ollama address
in Settings → Prompt writer (or OLLAMA_HOST), like 192.168.1.20:11434 or http://ollama:11434.

It looks before it writes. Add your references first: the writer is shown your reference pictures,
two frames of each reference video, your first and last frames, and the last frame of a video you're
continuing (up to six images, labelled <Picture 1>, <Video 1> and so on), and writes what it sees.
It tells you what it will look at, or that the model you chose can't see pictures.
OpenAI-compatible: any server that speaks the OpenAI API (/v1/chat/completions): llama.cpp's
llama-server, llama-swap, vLLM, LM Studio, LiteLLM, LocalAI, or OpenAI itself. Enter the address (like
localhost:8080; /v1 is added when there's no path) and the model name, and turn on It can see
pictures for a vision model. An API key and extra headers are optional and stored encrypted. It writes
exactly as the local writer does, asking for structured JSON (json_schema, falling back to json_object
or plain text for servers without it). IV Studio doesn't load or unload its models: with llama-swap, the
model name picks what it loads. Agents using the MCP server pass the key as IV_STUDIO_LLM_KEY (or
OPENAI_API_KEY) and headers as IV_STUDIO_LLM_HEADERS.
Claude Code, Codex or Hermes: if you have one installed and signed in, it writes with your own login,
headless and with no tools but reading your reference pictures (Hermes, from 1.1.7, on the model you picked in
Hermes; it sees your pictures when that model can). It gets MiniMax's own H3 prompt guide
(downloaded from MiniMax on first use, not shipped) and the full directing method.
MiniMax Context IR: MiniMax's own prompt rewriter through ComfyUI, paid through your Comfy.org
account (about $0.02 to $0.11 a prompt). Your API key is stored encrypted on your PC.

Projects: a short film in six steps

Projects turns an idea into a finished short film, one step at a time, with a storyboard you can see.

Idea. What it's about, the format (16:9, vertical or square) and the length.
Story. Write it for me drafts a script from the idea, cast and length with your prompt writer (Settings → Prompt writer:
the local model, Claude Code, Codex, Hermes or an OpenAI-compatible server), or
write it in plain words, or bring a .txt or .fountain screenplay.
Cast and places. Speakers are played by your characters. Someone you made up gets a face drawn from
their description, or from your own picture, and keeps it in every shot. Places get a description and a
picture (drawn for you if you have none), plus More angles (a quarter turn, the side and a wide view,
drawn with the QMA camera-angles LoRA for Qwen Image 2.1). Pick the film's look from six styles or your
own words and LoRAs, add Cinematography (film stocks, lenses, lighting setups, camera rigs, moods: 165
ready phrases, each shown as a picture), and make a Lookbook: name a film, director or cinematographer, or describe the look,
and your writer turns it into look language with no names in it. All of it goes into every shot.
Storyboard. The writer plans each scene into 3 to 10 second shots with staging and where each one
ends, writes each H3 prompt, and every panel gets a sketch that keeps your characters' faces and the
room. Open a panel to change the action, camera, speaker, line or length, draw it on the staging board,
use a recording of the line, or carry straight on from the shot before in one take. Play the
animatic shows the whole film as sketches with the lines spoken.
- Coverage. Cover a scene with one of twelve coverage plans: full coverage, dialogue coverage (a
  wide master, a matched over-the-shoulder pair, a matched close-up pair and a listening reaction),
  motion, extreme action, an establishing sequence, surveillance, an entrance, parallel action, a dance,
  angle coverage, an orbit or story beats. The planner then makes exactly those shots, each named by its
  slot.
- Check continuity reads a scene's shots like a script supervisor (looks and wardrobe, props, light,
  time of day, weather, screen sides and eyelines, and how each shot ends against the next) and gives a
  concrete fix for each mismatch.
- Prompt tools on every shot: a director's note in plain words ("make it colder, she shouldn't
  smile"), Tighten, Enrich, Distill, Stronger shot, New angle and 3 variants to
  compare. Spoken lines, labels and sections are kept, and every earlier prompt stays in the shot's
  history to go back to.
- Shot list: save it as a Markdown document or as CSV for a spreadsheet.
The setups and coverage plans come from Slate, the prompt studio for AI filmmaking, by Sam Wasserman
(wassermanproductions.com, Apache-2.0), through Imagination Studio by keys + Wasserman. They ship
unchanged in resources/slate/ with Slate's notice.
Film it. Every panel is filmed from its sketch (2 or 3 takes), in References mode with the place's
picture and the faces on screen, so the room and the people hold. Lines are spoken first in each
character's own voice and read back with Whisper until the words match. Pick a take, and it finishes at
2 MP.
Watch. The takes play in order; Save the film as a video exports the rough cut, and Fine-tune in
the editor opens it in Edit.

Images

Create switches between Video (a shot) and Image at the top. Image makes pictures on your PC with
the image models you have in ComfyUI:

Family	Generate	Edit	Licence
Flux.2 Klein 9B	✓	✓ several reference pictures	FLUX Non-Commercial (9B)
Flux 3 Klein	✓	✓	FLUX Kommunity
Krea 2	✓	✓ with the Krea 2 Identity Edit LoRA and the comfyui-krea2edit nodes	see Krea 2's model page
Moss	✓	needs a newer ComfyUI	Apache-2.0
Qwen Image 2.1	✓	✓ up to three references	Apache-2.0
Z-Image Turbo	✓	✗	Apache-2.0
SDXL (and any SDXL fine-tune: Juggernaut, RealVis, Pony, Illustrious…)	✓	✗	CreativeML Open RAIL++-M (base); fine-tunes have their own
MiniMax H3 (a still)	✓ with your H3 character LoRAs	✗	MiniMax H3 Community License

Getting a model (1.1.3). A family you don't have yet says not set up. Click it to see the files it needs,
which of them you already have, and one button that downloads the rest into ComfyUI (pinned versions, each
checked against its published SHA-256). The same files are under Settings → Setup → Optional tools → Picture
models. Your own builds and fine-tunes count: a Moss or Krea 2 fine-tune in its own folder makes that family
ready. Flux.2 Klein 9B and Flux 3 Klein are gated: accept their licence on Hugging Face, then save a Read token
in that panel or on Setup (or log in with hf auth login).

SDXL checkpoints go in ComfyUI's models/checkpoints; Setup offers SDXL base 1.0 as an optional download. An
H3 still is one frame from the middle of a very short H3 render (8 steps with the turbo LoRA), so it has the
look of your shots and can star your trained characters; it uses the H3 models you already have.

Pick Generate or Edit; Edit lists only the families that can. Pick a family and one of its
checkpoints (your fine-tunes are there too), add that family's LoRAs from the same picker as Video's, then
the size, how many (1 to 4) and a seed. Each graph follows ComfyUI's official template for its model.

Finished pictures appear under the form and in the Library. Open one to Use in a shot (it becomes the
shot's first frame), Edit this, Add to edit, Make more like this or Show file. Agents use
list_image_models and make_image.

Characters

A character is a trained MiniMax H3 LoRA plus its trigger word, or a RefMod; write {char} in the prompt
where the character goes. On the Characters page:

Your characters: portrait cards; click one for its trigger, strength, LoRA or RefMod (checked against
ComfyUI) and voice. Add from a LoRA or RefMod adds one you already have.
Ready-made: IV Studio's own invented characters, Tess Marlow and Theo Lindqvist. Download one and
its LoRA (about 150 MB, from this repo's characters-v1 release, checked against its SHA-256) goes into
ComfyUI, and their bundled voice comes with them.
Make a RefMod: up to six pictures (and a voice clip) become a RefMod character on your PC in seconds,
with no training (ComfyUI-MiniMaxH3Mod's RefMod extractor). It works in Create's shots.
Train a LoRA, the method that made Tess, for the best likeness:
1. Design: describe the face; Z-Image Turbo renders four candidates on your PC. The one you pick is
  rendered in eight scenes on the same seed. Or add your own pictures, only of an invented person or
  someone who agreed.
2. Clips: H3 renders talking clips from the stills and the character's voice, or you add your own.
  Everything is fitted to Fizgig's rules (24 fps, 17n+5 frames, 32 kHz stereo).
3. Train: with Fizgig, where you choose:
  - This PC, free: on your own NVIDIA card with 16 GB or more. A Standard run is a few hours; when
    Check my PC finds a card that can, the page says Free on your PC and the Train step starts there.
  - RunPod or Modal: your own account, by the second, with the minutes and dollars shown first. Theo
    took 56 minutes and about $4.10 on an H200. On RunPod, IV Studio makes an 80 GB storage volume (kept
    for later runs) and a training pod that deletes itself when the run is done.
  The recipe: Quick, Standard or Thorough; look and voice, or look only; normal or fine detail; the person
  or their whole look; how many checkpoints; and, under Advanced, rank, epochs, seed, caption dropout,
  schedule and base precision. A spending cap stops a cloud run at your budget and keeps what it saved.
  Train more carries a finished run on from its last saved state.
4. Pick: checkpoints are rendered side by side on one seed, and Score them for likeness marks the
  closest; the one you choose becomes a character, keeping the voice its clips were made with.
Likeness: with InsightFace's face models (a one-time 289 MB download; non-commercial licence), IV
Studio scores every still and clip against the picked face and can keep the closest of each scene.

LoRAs

LoRAs (on the Visual side) changes how H3 looks and moves: styles, motion, camera, speed.

Discover searches Civitai's MiniMax H3 LoRAs: tall cards whose preview fills
the card and plays on hover, filters (style, concept, cinematic, motion, camera, effect, speed, tool) and
sorting. Open one for a pop-up with a gallery of up to 16 examples (click one for a full-size view with
its generation prompt and settings, when Civitai has them), the maker, versions, trigger words to copy,
the licence (whether commercial use is allowed), tags and the maker's notes, then Download: the file
goes into ComfyUI's loras\Civitai\<model>\ folder, checked against the
SHA-256 Civitai publishes. Downloading needs your own free Civitai API key (Settings → Civitai; it's
stored encrypted on your PC). Searching doesn't.
Yours shows the LoRAs already in ComfyUI, grouped by folder, H3 ones first. Find on Civitai looks
each one up by its file's SHA-256 for its name, trigger words and a preview. Star your favourites.
Create's LoRA tray shows the LoRAs in a shot as pictures with a strength slider; + Add LoRA opens
the same browser. A LoRA's trigger words go into the prompt when you add it.
Mature LoRAs are off. Settings → Civitai → Show mature LoRAs (18+) switches to civitai.red, with
previews blurred until clicked; a second box, Show mature previews without the blur, drops the blur.
LoRAs of real people and anything involving minors are never shown, in either mode.

Voices

Audio → Voices keeps a voice for every character, and Audio → Lines says anything in any of them.

Design a voice from a description ("a woman in her forties from Glasgow, low, warm and a little
gravelly"): three takes to choose from.
From a recording: a clean 10 to 20 seconds of one person talking. Only your own voice, an invented
one, or someone who agreed, and IV Studio asks you to confirm.
Speak it: in Create, a character with a voice gets a Speak it box. Type the line; it becomes
the shot's voice line, and the same words go into the prompt so H3 lip-syncs them.
Lines keeps every line you've spoken, ready to use in a shot or drop on an edit. A new seed gives a
new read of the same words.

The speaking runs on your PC in IV Studio's own Python environment (set up once, about 12 GB):
VoxCPM2 (OpenBMB, Apache-2.0) designs, clones and speaks, and
Whisper large-v3 turbo (OpenAI, MIT) reads the
words in a recording. A line takes about 4 seconds once the model is loaded. The voice engine is stopped
before every render, so it never shares the GPU with H3.

Music

Audio → Music makes songs and instrumentals on your PC through ComfyUI's own nodes:

MiniMax Music 3: full songs up to five minutes with the clearest sung words of the open models.
Describe the style and write lyrics with [Verse], [Chorus] and other section tags; the song's length
follows the lyrics. About a minute or two per song on an RTX 4090.
ACE-Step 1.5 Turbo: a song in seconds, with tempo, key, time signature and language. Best for drafts,
background music and instrumentals.

Every song is kept in IV Studio's music folder. Add to edit puts it on A1 of a new or existing edit;
Use in a shot makes it the voice line, so a character sings it (trim it to the shot first).

Sequences

Several shots that play as one continuous video.

One take: every shot renders together in one seamless take through
H3 Multishot's sampler, so there are no cuts,
colour shifts or breaks in the sound.
Shot by shot: hunt and pick each shot in turn; each carries on from the end of the one before.
Retake: mark a stretch of any version and redo the picture and sound, the picture only, or the sound
only. Everything around it stays.
Render out: optional colour smoothing across the joins, then DLSS with RTX 2× and frame generation.

Shots are 10 or 15 seconds, the lengths H3 Multishot is tested at.

Edit

An editor laid out like Adobe Premiere for cutting shots, sequences, songs and lines together. Full
screen (or F11) hides IV Studio's sidebar and gives the whole screen to the edit; Esc brings it back.

Panels. Project bin, Source monitor, Program monitor and Effect controls on top, the timeline below.
Drag any border to resize; the layout is remembered.
Project bin. Thumbnails, search, and Video, Pictures and Sound bins. Its Library tab holds every IV
Studio render, song and line, one click from the edit.
Source monitor. Click a bin item to open it, mark In (I) and Out (O), then press , to insert it at the
playhead (everything after moves right) or . to overwrite. Or drag the picture onto a track.
Timeline. As many video and audio tracks as you need (a higher V track plays over a lower one), plus
titles. Each track can be targeted (click its name), hidden, locked, muted or soloed. Video clips show a
filmstrip and their sound's waveform; sound clips show a waveform. Markers sit on the ruler.
Tools. Select (V), Razor (C, click to cut; the cursor becomes a red katana), Ripple trim (B, the rest of the track follows) and Rolling
edit (N, moves a cut between two clips). Each clip has fades and volume; titles have their own text,
size, colour, position and box.
Export MP4 renders the timeline at 1080p, 4K, vertical or square, 24/30/60 fps, into
Videos\IV Studio\Edits. Hidden tracks stay out of the picture and muted ones out of the sound.

Keys, as in Premiere: Space play · J K L shuttle (press again to go faster) · ← → one frame (Shift: one
second) · ↑ ↓ previous / next edit · I O mark · , insert · . overwrite · S split · Ctrl+K split every track ·
Q W ripple trim to the playhead · Delete, Shift+Delete ripple delete · M marker, Shift+M next marker ·
\ fit · = − zoom · Ctrl+Z / Ctrl+Y · F11 full screen, Esc back.

Make new footage on the timeline. Right-click a clip or a gap:

Extend… carries a clip on 5, 10 or 15 seconds from its last frames.
Bridge to the next clip… makes a new shot from one clip's last frame to the next one's first.
Fill this gap… does the same across a gap.
Retake this stretch… redoes part of a clip: picture and sound, picture only, or sound only.

Each one starts from the clip's own shot when IV Studio made it (its character, LoRAs and look), and
Write it for me looks at the frames it starts from. A striped placeholder shows where the footage will
land and how far it has got. Bridges and fills offer three takes to pick from; the result lands as a
normal clip you can undo.

More on the right-click menu:

Cut out… keeps only who or what you name ("the bluebird", "the man in the hat"). SAM 3.1 tracks it,
and the clip comes back with see-through edges on the track above, sound kept.
Reframe to 9:16 / 1:1 / 4:5 / 16:9… fits a shot or picture into the edit's shape and draws what lies
beyond its edges: H3's masked video edit for clips (the original picture and sound are kept), Qwen Image
2.1 with an outpaint LoRA for pictures.

Loudness. Exports come out at Online loudness (about -14 LUFS, the default), TV (-23) or as mixed, under
a -2 dBFS peak limit.

With an agent. Your agent edits the same timeline through IV Studio's MCP server. Its changes appear
at once with a strip saying what it did and a one-click Undo, and tracks you lock are off limits to it.
It can make new footage on the timeline too.

Library

The Library button above the Visual | Audio switch opens everything you've made: shot finals and
previews, sequences, Replace runs, pictures, Edit exports, songs and spoken lines, newest first. A shot's
previews and finals are one tile with an "N versions" badge; click it to watch every version in a large
viewer and step through them with the arrow keys. Everything sits under the day it was made (older work by
month), filtered by type (videos, images, music, voice lines), when, where it came from (Shots, Sequences,
Replace, Edit exports, Images, Music, Voice lines) and, for shots, their kind, who's in them and their shape;
sort by newest, oldest, name, most versions or longest. Search also finds words in the prompts renders were
made from. Tiles show each render's size, length and model.

Open a tile to watch its versions, listed as rows with their full names. How it was made shows the model,
kind of shot, size, length and frame rate, character, LoRAs and RefMods, seed and steps, which take it is, how
it was finished and enhanced, the render time and what it cost, with the prompt it came from (Copy,
Show all). Then add any of them to an edit or use a sound as a shot's voice line.

Above it, the Queue shows what ComfyUI is rendering and what waits, on every page: each render's name,
its step and time, a click to open its page, and × to cancel one of IV Studio's own. Renders an agent
started show as rendering with their time, without the step count.

Simple words

New to AI video? Check my PC asks how much you know about making AI video and pictures, and Settings →
Words has the same switch. With Simple words on, IV Studio explains itself in everyday words: a LoRA is a
"trained look", a seed a "variation number", DLSS a "quick sharpen", an upscale an "enlargement", a prompt a
"description", and the main help lines are rewritten in plain language. Only the words change: every page,
control and setting stays the same, and names, prompts and file paths are never touched.

Accessibility

Settings → Accessibility makes IV Studio easier to read and use, and each change applies at once:
Colour (Red, Blue, Purple, Orange or Dark: buttons, highlights and the artwork follow),
Text size (Normal, Large, Larger: everything scales, like browser zoom), Higher contrast text (grey
secondary text close to white, stronger lines; also offered on Check my PC's first page), Reduce motion
(no animations, and the Workflows stages stand still; your system's own setting is followed too), Clear
focus outlines (a thick yellow ring on whatever the keyboard is on) and Underline links.

Incognito

The Incognito switch in the title bar restarts IV Studio into a separate, private space, and the eye
closes. Shots, sequences, edits, songs, voices and lines, characters, trainings, Shinobi films, sizzles and
exports made there are kept apart (in %APPDATA%\IV Studio\incognito, exports in its videos folder) and
never show in the normal app; nothing normal shows there. Your characters and saved voices come along, and
settings, keys and models are shared. ComfyUI saves Incognito renders into its output\IV Private folder
(uploads into input\IV Private), and leaving Incognito clears them from ComfyUI's history. Switch back
the same way.

Settings → Folders → Finished videos picks where exported films and edits, and cloud renders brought to
this PC, go: any folder on any drive (Videos\IV Studio unless you change it; Check my PC asks too).
It also picks where in ComfyUI's output folder renders save, one folder for normal renders
(straight into output unless you name one) and one for Incognito (IV Private unless you change it). They
can't be the same folder, and they always stay inside ComfyUI's folders.

Incognito keeps things out of IV Studio's normal views. It isn't encryption: the files are on your PC, in
the folders above, and LoRAs trained in Incognito land in ComfyUI's LoRAs folder like any other.

Workflows

Seven tools for what comes after a render: Sequences, Sizzle, Effects, Reel, Shinobi, Replace and Rebuild. Each
opens on its own animated stage with its main bar centered (a wall of fire for Sizzle, a beat for Reel, running
film for Sequences, blade cuts for Shinobi, a scanning swap for Replace), then works in numbered steps; the
stages pause when they're off screen and stay still if your system asks for reduced motion. Sequences is
described above, and Effects in what's new in 1.1.

Rebuild takes a video you made apart and remakes it: IV Studio reads it shot by shot, your agent (Claude Code
or Codex) plans how each shot is remade, you approve the plan and the frames, H3 renders the new shots on this PC,
and a first cut lands in Edit beside the original.

Sizzle

Put in a web address and get a short promo film for it.

Capture. IV Studio opens the site in your PC's Edge or Chrome (hidden) and saves what people see: a
screenshot of every screen down the page, the phone layout and the logo, plus the site's words, colours,
fonts and pictures. Cookie banners are hidden, never accepted.
Storyboard. Your agent writes the scenes from the capture: the site's own screens, H3 shots,
Z-Image pictures, titles from the site's copy, music and an optional voice-over. Nothing renders until you
approve it.
Pieces. H3 shots (previewed, then finished at 1 MP), pictures, music and the voice-over are made on
your PC: about 3-5 minutes per H3 shot.
The film. Cut it for me puts everything in a Shinobi project and cuts a first version: cuts on
the music's beat, the site's screens with slow push-ins, two-tone titles in its colours, and an end card
on the logo, the main button's words and the web address, ending where the music ends. It renders in
about 2-3 minutes. Your agent can then refine it in Shinobi (the site's own fonts, pans, effects).

The Sizzle page shows the capture, the storyboard with each scene's piece, and the finished film, with a
message to paste into your agent. Agents use the sizzle_* tools.

Reel

Pictures or clips plus a song become a new edit cut to the beat: a whole number of beats per picture (calm,
normal or fast), cuts counted in bars from the first beat, a slow push, drift or pull on each picture, in
16:9, 9:16 or square. The beat is found with the bundled FFmpeg alone, within a few milliseconds on click
tracks.

Replace

Puts your characters into a video, or changes the scene around the people in it, while the camera, cuts,
motion and soundtrack stay:

Swap people: one or two characters (your presets, Tess and Theo, or any picture) take the place of
whoever you describe: face, hair, body and clothes. A preset's LoRA can join the run.
Change the scene: in plain words: remove or add things, or change the weather, light, setting or style
("a snowy winter evening", "remove the wind turbines").
Finish: redraw small faces at close-up size, or soften H3's look.

It runs on Viggle Remake, two LoRAs on MiniMax H3 (swap and
edit), up to 14 s at a time. IV Studio does the parts Viggle leaves to a person:

it puts the source on Remake's canvas (24 fps, its own shape, the frame count H3 needs) and finds the shots;
it paints a frame of each shot the way the result should look, with Qwen Image 2.1 edit on your PC. A swap's
frames are checked with the face model (nobody added, and it looks like your character) and painted again
when they fail; then you check them (paint one again, use your own picture, or leave one out);
your prompt writer (local, Claude Code or Codex) writes Remake's six-section prompt from the source and the
painted frames, following Viggle's own guide, and IV Studio checks it;
Remake renders 1 to 3 takes (about 2.5 minutes for 5 s on an RTX 4090), trimmed back to the source.

In the editor, right-click a clip → Remake this clip…. Setup downloads the two Remake files (21 GB each)
and installs Viggle's nodes. Powered by Viggle Remake; the weights are under the MiniMax H3 Community License.

The old engine, OpenJutsu's mask workflows (MIT, commit 47c3724), is
still there behind Use the old engine at the foot of the page for one more version; it runs Finish.

Shinobi

IV Studio's agent-driven editor. Where Edit is cut by hand, Shinobi lets an agent (or a one-click recipe
from 12 styles) author the whole edit in code: cuts on the beat, speed ramps, punch-ins, typography, masks,
captions from speech and sound cues. It renders on your PC with the Edge or Chrome you already have.
Shinobi's engine is OpenKatana (Apache-2.0), installed unmodified;
Shinobi isn't affiliated with OpenKatana or with Higgsfield, whose Katana inspired it.

Help on every page

How it works: the Create tour lighting up What happens, with its card and Next

Every page has a How it works button by its title (and in the editor's top bar). Click it and the page
walks you through itself: each part lights up in turn with a line about it, and Next, Back, the arrow keys or
Esc move through it. The button glows until you've taken that page's tour once. Steps follow the page as it
is, so Create's tour includes the reference step only in the modes that take references.

The welcome wizard (Check my PC) ends by offering a full tour: about two minutes across the whole app,
ending on the How it works button. Take it again from Settings → About, where the PDF guide is too. After an
update, What's new shows on the Dashboard once.

How IV Studio uses ComfyUI

IV Studio is a front end for your own ComfyUI; every picture, video and song is a ComfyUI job.

The API. It queues graphs with ComfyUI's HTTP API (/prompt, /upload/image, /history, /view,
/object_info, /free) and follows progress on its websocket, including KJNodes' live step previews.
Before a render it checks /object_info for every node and model the graph needs.
Graphs built in code. Each workflow is assembled per job from the user's choices in src/service/
(studio.js for shots, images.js, music.js, reframe.js, editgen.js, productions.js and more),
following ComfyUI's official templates for each model. The H3 hunting graph is checked against its
ComfyUI GUI export in 32 switch states (test/graph-parity.mjs).
Workflow files. resources/workflows/iv_studio_api.json is the full IV Studio H3 workflow in API
form (with iv_studio_api.groups.json, its switch groups), and resources/workflows/replace/ holds the
Replace's old-engine workflows from OpenJutsu. Load them in ComfyUI to see the graphs IV Studio drives.
Remake's graph is Viggle's own workflow (src/service/remake.js builds it per take).
Node packs. Settings → Setup → Install missing node packs adds every pack IV Studio uses to your
ComfyUI's custom_nodes in one go: its own small pack (resources/comfy-nodes/ComfyUI-IV-Studio) is
copied, the others are downloaded at the versions below (the list lives in resources/models.json), and
their requirements are installed with ComfyUI's own Python (never the torch family, which ComfyUI already
has). A folder that's already there is left alone. Restart ComfyUI afterwards. Agents can do the same
with install_node_packs. Replace and Cut out also use SAM 3, which recent ComfyUI has built in.
Updates. Settings → Setup → Updates checks ComfyUI against its newest stable release and counts
the custom node packs with updates. Update everything saves a ComfyUI-Manager snapshot, updates every
pack (Manager included) and ComfyUI through Manager, then restarts ComfyUI and waits for it; nothing runs
while ComfyUI is rendering. Undo last update restores the snapshot, and switches back on the plain-folder
packs (IV Studio's own among them) that Manager's restore switches off. It needs ComfyUI-Manager (Setup
offers to install it) and a ComfyUI on this PC; under the ComfyUI Desktop app, which updates ComfyUI itself,
it updates the node packs. Agents: check_comfy_updates, update_comfy, undo_comfy_update.

The latent upscaler. ComfyUI 0.44 and newer have the H3 latent upscaler built in; there, finishes use
ComfyUI's own (same model and size) and the LBH node pack below isn't needed.

Pack	Used for	Version
ComfyUI-MiniMaxH3Mod (RefMods)	Shots	Comfy Registry 0.2.9
Fantastic MiniMaxH3 PromptBuilder (media loader)	Shots	commit `22e719f`
MiniMax H3 latent upscaler	Shots (ComfyUI before 0.44)	commit `40316cf`
VideoHelperSuite	Shots	commit `4d907be`
DLSS5 Enhancer v15 (Strix-3D)	Shots (Windows only)	Comfy Registry 1.3.1
rgthree-comfy	Shots	commit `d7297b4`
Impact Pack	Shots	Comfy Registry 8.28.3
Pixaroma (LoRA loader)	Shots	commit `30d4d94`
Easy-Use (seed)	Shots	commit `b0b47ca`
comfyui_essentials (math)	Shots	Comfy Registry 1.1.0
KJNodes	Shots	commit `37ca7af`
IV Studio nodes (latent cache, shot list)	Shots	bundled with IV Studio
H3 Multishot	Sequences	Comfy Registry 2.7.2
MiniMax H3 Context Loop	Replace and Reframe	Comfy Registry 0.7.3
H3 One-Click Swap	Replace	commit `e44a56e`
H3 Face Refine	Replace	commit `d8521d1`
Image Detail Contrast	Replace	commit `485a5e2`
Orion4D FXMax	Replace	commit `7048e84`
Spectrum for MiniMax H3	Replace	commit `806fe49`
SeedVR2 Video Upscaler	SeedVR2 finishing	Comfy Registry 2.5.24
Frame Interpolation (RIFE)	Finals on Linux (Linux only)	Comfy Registry 1.0.11

Memory. One model family at a time (below): IV Studio calls /free when the next job needs a
different family.

Models and the GPU

Free (next to the GPU and RAM readout in the title bar) unloads everything from the GPU and RAM now:
ComfyUI's models, IV Studio's voice engine and the prompt writer's Ollama model. A render that's running
keeps its memory until it finishes.
Settings → Models picks the main H3 model for each workflow from the ones your ComfyUI has, so a
fine-tune that suits realism or animation can be used where it fits.
One model family at a time. Before every job, IV Studio stops its own voice engine and, when the job
before used a different family (H3 video, music or design pictures), asks ComfyUI to unload first. Jobs
of the same family keep their models loaded.

Cloud

No strong graphics card, or want your PC back while it renders? The Cloud page renders on a GPU in your
own RunPod or Modal account instead: your account, billed to
you by them. Or use Comfy Cloud, Comfy's own hosted ComfyUI. Pick one on the
page, or press Guided setup for a few steps: pick a service, connect your account, choose a GPU and when it
shuts off, set it up, switch it on. Check my PC offers the same wizard when it recommends the cloud.

On Comfy Cloud

Comfy account. Paste a Comfy API key from platform.comfy.org
(the Comfy.org key you saved for the paid prompt writer works too). Comfy Cloud's API needs a paid Comfy
Cloud plan; runs are billed to your Comfy account, and nothing runs between renders.
What works there. Comfy Cloud runs ComfyUI's own nodes and a fixed set of node packs that Comfy
maintains, with its own model library; you can't install node packs there. So IV Studio asks it (free, no
render) and lists what works: pictures and music use ComfyUI's own nodes, while video shots need IV
Studio's node packs and run only if Comfy Cloud offers them. Check again any time; Comfy adds models and
packs.

On RunPod

RunPod account. Paste an API key from RunPod's Settings → API Keys page (read/write). It is checked
with RunPod, then stored encrypted on your PC, and used only to make your storage volume and to start and
stop your own pods.
GPU and shut-off. Pick an L40S or RTX 6000 Ada (48 GB), A100 or H100 (80 GB), RTX 5090 (32 GB) or RTX 4090 (24 GB). The
page shows RunPod's live hourly price and whether each is free right now; billing is by the second.
Set up. IV Studio makes a 100 GB network volume in a RunPod data center that has your GPU and starts a
pod from RunPod's official PyTorch image with this version's resources/cloud folder. The first pod installs
ComfyUI and the same pinned node packs as Modal onto the volume, downloads the H3 models (each checked
against its published SHA-256) and receives your character LoRAs, so later pods start in a few minutes. When
a new IV Studio changes the cloud setup, the next pod installs it again by itself (your models stay).

RunPod pods don't sleep by themselves, so the pod's gate (resources/cloud/gate.py, which also asks for IV
Studio's token on every request) deletes the pod after your idle time with an empty queue. The volume keeps
ComfyUI, the models and every render. The next render starts a new pod; if your GPU is taken in that data
center, IV Studio starts the next H3-ready one that's free. The volume is billed by RunPod while it exists
(about $7 a month for 100 GB); delete it on RunPod's Storage page when you're done with the cloud.

On Modal

Cloud tools. IV Studio installs a private copy of Modal's client in its own folder (about 100 MB).
Modal account. Sign in on Modal's own page; signing up is free.
GPU and shut-off. Pick an L40S (48 GB, recommended), A100 or H100 (80 GB), and how long it waits
with nothing to do before it switches off (3, 5, 10 or 20 minutes). The Cloud page shows each GPU's
hourly price; billing is by the second, only while the GPU is on.
Set up. IV Studio deploys its ComfyUI app to your account and copies the H3 models into a Modal
volume (about 60 GB) once. After an IV Studio update that changes the cloud setup, press Set up again
once (your models stay).

Either way

Then switch This PC / Cloud. While you're on the cloud, the title bar shows GPU on/off · time · $:
the GPU time and estimated cost since IV Studio opened, at the chosen GPU's list price. Click it for the
Cloud page, where Shut down GPU stops it at once instead of waiting for the idle shut-off. The next
render wakes it again (about 1 to 3 minutes on Modal, 2 to 5 on RunPod). The Cloud page estimates a 10-second 2 MP shot at roughly 3 to
5 minutes of GPU, about $0.13 on an L40S.

What runs on RunPod and Modal: nearly everything. Create's video shots and Sequences, pictures (every
family, Generate and Edit) and the picture tools, Projects sketches, designing a character's stills and clips,
Music, Sounds, Cut out, Reframe, Replace (Swap people and Change the scene, on Viggle Remake), Effects on clips
and photos, Retake in the editor, the SeedVR2 finish and Enhance, and Voices (Lines, Speak it, and Whisper
reading lines back). Finals use a 2× AI upscale (Real-ESRGAN)
and RIFE frame interpolation (48, 72 or 120 fps), or SeedVR2, in place of DLSS 5, which needs an NVIDIA card on
Windows. The cloud runs ComfyUI v0.44.1 with an exactly pinned torch, and takes its node packs from the same
pinned list Setup installs on a PC.

Models on first use. Setup puts only H3's models in your cloud storage. Everything else brings its own
models the first time you use it there: IV Studio lists them with their size and asks before downloading
(on Modal on a CPU, a few cents; on RunPod in the GPU you're already using). Each file is checked against its
published SHA-256 and kept for next time. Picture families use one pinned model each in the cloud (Flux.2
Klein 9B, Flux 3 Klein, Krea 2 Turbo, Moss, Qwen Image 2.1, Z-Image Turbo and SDXL base); on a PC your own
files are used. If a download won't fit on your RunPod volume, IV Studio says how much bigger it must be and
what that adds a month, and grows it only if you agree (RunPod volumes can grow, never shrink; about $0.07 per
GB a month). Modal's storage is free up to 1 TiB.

Gated models. Flux.2 Klein 9B, Flux 3 Klein and the LTX effects (Dub, Clean plate, Day to night, Add
water, SDR to HDR, Transformations and Transitions) are behind a licence on Hugging Face. To use them in the
cloud, accept each licence on its Hugging Face page, then paste a Read token on the Cloud page (stored
encrypted; used only for those downloads). The same token serves downloads on your PC (Settings → Setup).

Your LoRAs and RefMods. A cloud render that uses a LoRA or RefMod from this PC uploads it first, once,
checked by SHA-256 (from your local ComfyUI's folders, or IV Studio's own models folder on a PC without
ComfyUI).

On Comfy Cloud the check above decides it: Images and Music run there when Comfy Cloud has their models.

Voices in the cloud. IV Studio's own voice engine (VoxCPM2 and Whisper, the same as on a PC) runs beside
ComfyUI on your cloud GPU. Press Set up voices in the cloud once on Audio > Voices: about 12 GB goes into your
cloud storage (on Modal its Python is part of the cloud image and the models download on a CPU; on RunPod a GPU
installs both onto your volume in 10 to 20 minutes, about $0.30). Each line then runs on the cloud GPU, billed by
the second; the engine stops as soon as a render is queued, and after 10 idle minutes.

What still needs this PC: live step pictures while a shot renders (they need a
connection held open, which would keep the GPU billing), Replace's old engine, the H3 pass of the free
Seedance finish, and New angle. Training a character runs on your PC, RunPod or Modal, whichever you pick.

Agents (MCP)

IV Studio includes a local MCP server (stdio) with 214 tools. It uses the same projects, characters, voices
and songs as the window, so Claude Code or any MCP client can render shots and sequences, make pictures,
speak lines, make music, cut them together and export. Settings → Agents has the command for your
install, a message to paste into any agent, and a button that adds the IV Studio skill to Claude Code.

Start menu. When a chat opens without a request, the agent offers IV Studio's menu as one
multiple-choice question: a shot, a picture, a longer scene, a spoken line, music, an edit, a website film,
a swap, a new character or a tour. Each choice comes with what to do next (get_menu). The same choices,
plus start, are MCP prompts, so Claude Code lists them as slash commands (/mcp__iv-studio__…).

For a development checkout:

claude mcp add iv-studio -e ELECTRON_RUN_AS_NODE=1 -- "<checkout>\node_modules\electron\dist\electron.exe" "<checkout>\out\main\mcp.js"

Agents learn how to use it from resources/agent-guide.md: the server's instructions on connect, the
get_guide tool, and the Claude Code skill (skills/iv-studio/SKILL.md).

All tools

Guide and start menu: get_guide, get_menu.
Images: list_image_models, make_image, list_images, get_image, cancel_image, delete_image, last_frame.
Activity: list_activity.
This PC: scan_hardware (what Check my PC sees and recommends), free_memory.
Shots: check_setup, install_node_packs, check_comfy_updates, update_comfy, undo_comfy_update, list_characters, list_projects, plan_render, check_prompt,
list_samplers, generate_previews, finish_candidate, enhance_final, rank_takes, use_clean_stretch, review_take, make_previs, soften_mouth.
Cast and places: list_cast, save_place.
Projects (short films): list_films, create_film, get_film, update_film, set_film_script,
write_film_script, break_down_film, sketch_film_shots, edit_film_shot, write_film_shot_prompt,
speak_film_line, use_film_recording, draw_film_face, use_film_face, draw_place_angles,
make_animatic, render_film_shot, render_film_scene, finish_film_shot, build_rough_cut.
Sequences: list_sequences, get_sequence, save_sequence, render_take, hunt_shot,
finish_shot, retake, render_out.
Prompt writing: write_prompt, write_shots_from_story.
Jobs: get_job, cancel_job, get_outputs, make_dailies (an overnight batch's reel and notes).
References: prepare_media (crop a picture, trim and crop a video or sound, or fit it to H3's sizes).
Library: list_renders, rename_project, delete_project, delete_sequence.
Recommended LoRAs and turnarounds: recommended_loras, download_recommended_lora, make_turnaround, turnaround_status.
Edit: list_edits, get_edit, create_edit, set_edit_settings, delete_edit, add_media,
edit_timeline (add, insert, overwrite, move, trim, split, ripple delete, fades, volume, titles, tracks
and their hide/lock/mute/solo, markers, all checked before anything changes), grab_frame, export_edit,
get_edit_export, cancel_edit_export, export_edit_for_editors (Premiere, Resolve, Final Cut).
New footage on the timeline: generate_in_edit (extend, bridge, fill or retake), list_generations,
pick_generation, cancel_generation, cut_out, reframe, make_reel.
Voices: voice_status, setup_voices, list_voices, design_voice, create_voice, update_voice,
delete_voice, speak_line, list_lines, delete_line, release_voice_engine.
Music: music_status, make_music, list_music, get_music, cancel_music, rename_music,
delete_music.
LoRAs: search_loras, get_lora, download_lora (needs CIVITAI_API_KEY in the server's
environment), list_my_loras, identify_lora.
Characters: list_ready_characters, install_ready_character, save_character, delete_character,
list_choices (LoRAs and RefMods), make_refmod_character.
Character training: list_character_trainings, get_character_training,
create_character_training, update_character_training, design_faces, pick_face, design_scenes,
pick_still, make_training_clips, set_training_item, add_training_files, rank_likeness,
estimate_training and start_training (on this PC, RunPod or Modal, with the recipe options and more to
carry a run on; a paid run needs the person's OK and confirm_cost), stop_training, test_checkpoints,
score_checkpoints, adopt_checkpoint.
Replace: replace_check, replace_video, replace_history, remake_paint, remake_use_picture, remake_render.
Shinobi: shinobi_status, shinobi_setup, shinobi_tools, shinobi, shinobi_project.
Sizzle: sizzle_capture, view_sizzle_capture, list_sizzles, get_sizzle, save_sizzle_plan,
approve_sizzle_plan, make_sizzle_assets, sizzle_first_cut, sizzle_to_shinobi, delete_sizzle.
Models: list_h3_models, set_workflow_model.

An edit open in the window reloads when an agent changes it; a strip says what changed, with a one-click
Undo, and tracks the person locks refuse the agent.
The MCP server always renders on the local ComfyUI.

Use it well

Terms. Using MiniMax H3 through IV Studio means agreeing to the IV Studio terms of use: real people
only with their consent, nothing involving minors, say it's AI when you post it, and the rest of MiniMax's use
restrictions. How IV Studio guards against misuse, and what it doesn't do yet, is in
docs/SAFEGUARDS.md.
Report misuse. Seen something made with IV Studio that breaks the terms, or uses you or someone else without
consent? Email [email protected] (or Settings → About → Report
misuse). Every report is read and acted on.
Model licences. IV Studio doesn't bundle any model weights; each one downloads from its own host under
its own licence, linked from Setup. Read them before commercial work. In particular:
- MiniMax H3 is under the MiniMax H3 Community Licence, which covers use worldwide except the EU,
  the UK, South Korea and the USA. Use in those places needs a licence from MiniMax.
- MiniMax Music 3 is under the MiniMax-Music3 Community Licence: commercial use is allowed if the
  product shows "MiniMax-Music3", with MiniMax's written permission above $20M a year.
- ACE-Step 1.5, VoxCPM2 and Qwen3-VL are Apache-2.0; Whisper is MIT.
- InsightFace's face models (only for likeness scoring) are for non-commercial research use.
People. Characters and voices must be invented, your own, or of someone who agreed. Don't make a real
person say or do things without their consent, and don't clone a voice from a recording you found.
Music. Write your own lyrics and describe a style rather than imitating a named artist.

Develop

npm install
node node_modules/electron/install.js   # npm 11 skips Electron's install script
npm run build
"Start IV Studio (dev).bat"
npm run dist                           # unsigned NSIS installer in dist/

Tests are in test/:

graph-parity.mjs: the headless graph matches the ComfyUI GUI export for 32 switch states.
service-e2e.mjs: renders previews and a finish on a live ComfyUI.
mcp-smoke.mjs, mcp-tools.mjs, mcp-live.mjs: the MCP server over stdio, every editor, library,
model and character tool, and an agent's edit reloading in an open dev window.
mcp-voices.mjs: every voice tool against a set-up voice engine (pass its folder).
mcp-music.mjs: every music tool against the local ComfyUI (--minimax adds a MiniMax song).
mcp-loras.mjs: the LoRA tools against live Civitai and the local ComfyUI (downloads nothing).
civitai-download.mjs: Civitai downloads against a mock Civitai (checksum, missing key, real-person refusal).
edit-export.mjs: exports a five-clip edit with the bundled ffmpeg.
replace.mjs, shinobi.mjs, training-local.mjs, sequence-graphs.mjs: Replace, Shinobi, character
design and sequences on real hardware (the last one queues GPU renders).
media-prepare.mjs, models-download.mjs, prompt-writer.mjs: media trims, checked downloads and the
local prompt writer.
tours.mjs: every How it works tour exists, reads well, and points only at parts the window still has.
terms.mjs: which graphs count as MiniMax H3, the terms gate before H3 renders and training, and that
TERMS.md carries MiniMax's use restrictions.
hardware.mjs, training-recipe.mjs, training-runpod.mjs, sampler-choice.mjs, refmods.mjs,
news.mjs: Check my PC's advice, the training recipe and its spending cap, RunPod training against a mock,
the sampler choice, RefMod characters and What's new.

Licences

IV Studio is MIT-licensed (see LICENSE); the Infinite Variance eye artwork is Infinite Variance's
own and isn't covered by the MIT licence. Third-party notices are in
THIRD_PARTY_NOTICES.md. The bundled FFmpeg build is GPL.

IV Studio is Powered by MiniMax H3. MiniMax H3 is licensed under the MiniMax H3 Community License Agreement,
Copyright © 2026 MiniMax. All Rights Reserved. A copy of that licence ships in
resources/licenses/MiniMax-H3-LICENSE.txt, and using H3 through
IV Studio is subject to the terms of use.

Screenshots show Tess Marlow and Theo Lindqvist, invented IV Studio characters, and other shots, pictures and
songs made with IV Studio. The LoRA previews in the LoRAs screenshots are by their makers on Civitai.