H3 Studio - Comfyui - Lora Daddy -
A downloadable tool
Write it. Film it one clip at a time. Join it without a seam.
20 reference pictures on one rail. MiniMax's own node stops at 9. H3 Studio 2 does not. One person can carry their photos, their voice, and a stack of wardrobe outfits, and the rail still has room. The count on the panel is 20 pictures, 3 video clips, 9 audio files.
Watch the setup video on this page before you install. That video is the setup. Do not skip it.
LM Studio must be running as a server. Load a chat model, turn the local server on (it is http://127.0.0.1:1234 unless you changed it), and leave it on while you write. If the server is off, GENERATE does nothing. A thinking model (Qwen 3 and the like) needs Thinking / Reasoning turned off, or the script comes back empty.
H3 Studio 2 is a ComfyUI node pack for the MiniMax H3 video model. Type a sentence. Your own model, on your own PC, writes a timed H3 script: who is in it, every shot, the dialogue in each person's voice, the sound and the music. H3 renders it. The next clip carries on from the last one, picture and sound, with no visible join.
Four words in, a film script out. "Floyd waves saying hello!" comes back as his look, the wave, the wink, his line in his own voice, the sound and a score. On your PC. Nothing is sent out to write it.
Five tabs. One panel.
- T2V — text to video, with dialogue, accents and POV.
- I2V — your picture is the first frame, or give a last frame too and it films the journey between them.
- REF — your people. 20 pictures (MiniMax caps this at 9), 3 video clips, 9 audio files. Faces, clothes and voices held from shot to shot.
- MULTISHOT — a whole film, in one write or step by step.
- MUSIC VIDEO — drop in a song. It hears the lines, cuts the song into clips, and lip-syncs whoever you paint on the waveform.
Ref Bank and the wardrobe
People, places and things live on cards: up to 3 photos, a voice or a sound, and a clip. Drop a card into any tab. A private tab hides cards behind a password. 150 outfits ship in the wardrobe. Click one onto a person and that picture rides the rail with them, inside the 20.
FLOYD ships in the bank, with his voice. Lelu is the runway walk on the tour: each step is the next outfit.
The reference size toggle is 768 / 1024 / 2048. 2048 is a cap. It does not blow every still up to a 2048 short edge, which is how a few pictures run a GPU out of memory.
A free reference sheet maker is on this page as its own download. Drag images in, zoom, resize, set the aspect. It keeps the short side at 2048, which is what H3's reference node wants.
Step by step / extend
Make the film one clip at a time. Say what happens next, or press SUGGEST and it watches the last clip. Render, then keep it, re-roll it, or edit it. The writer sees the real last frame, so the same people stay in the same place in the same light. Start from nothing, from a clip you already made, or from a video you upload. JOIN writes the finished film.
On REF, the JOB dropdown says what a clip is for: copy the moves, replace the person, replace the head, change the outfit, change the background, or extend the video.
Music video
The song is the timeline. Paint who sings on the waveform. Each clip gets its exact slice, so the mouths land on the right words. Build it in one go or one slot at a time. Swap people and places between clips. The music stays in time.
Dials
- 26 styles (claymation, Ghibli, Minecraft, film noir, VHS…)
- 15 camera modes (selfie at arm's length, drone follow, tracking…)
- Colour grades, from black and white to teal and orange
- 569 film and TV worlds
- 33 accents, light, strong or heavy
- POV through someone's eyes
- Framing: no wide shots, so faces stay big
What you need
You need to know ComfyUI: how to download models, and how to use a subgraph for the first setup. Watch the setup video on this page. That is the install.
Updated weekly, on Fridays.
- Windows, ComfyUI, and MiniMax H3 (the model, the text encoder, and the VAEs). The video model is not in this download.
- LM Studio, with the local server switched on, and a chat model loaded. Made and tested with a Qwen 3 27B-class model. If the server is not running, the writer cannot start.
- A big GPU. Everything here was made on a 32 GB card. A 12B writer on a 16 GB card can work. H3 itself is the heavy part.
In the zip: the Studio, H3 Master (the renderer), H3 Encode, the Beta57 scheduler, and ComfyUI-H3-Motion-Context by NikoDemon80 (GPL-3.0, unchanged).
See it working: the interactive tour. Lelu's runway is the first thing on it. Click the wardrobe. GENERATE on that page does not render. It only shows the panel.
Every clip and screenshot was made with the tool. H3 Studio 2 is an independent tool. It is not made by or affiliated with MiniMax.
06/10/2026
added ninfer dropdown in settings . fixed shots. - more pov fixes
07-10-2026
NEW wardrobe. of clothing. without attribute transfer as its just the cloths without a body
| Updated | 1 day ago |
| Published | 5 days ago |
| Status | In development |
| Category | Tool |
| Author | Lora-Daddy |
| AI Disclosure | AI Assisted, Code, Graphics |
Purchase
In order to download this tool you must purchase it at or above the minimum price of £5 GBP. You will get access to the following files:






Leave a comment
Log in with itch.io to leave a comment.