MiniMax H3 Image
by MiniMax
Sixteen reference images keep one face across every frame.
Native 4K, drawn in about twenty seconds.

Key features
One face, many frames
Give it up to sixteen reference images, such as a character sheet, a location and the previous frame, and it keeps the person, the costume and the room consistent from shot to shot.
Native 1K, 2K and 4K
Every size is generated at full resolution rather than upscaled. 4K is about sixteen megapixels, enough to crop a close-up out of a wide frame.
Fast
A 2K frame with two reference images came back in about twenty-three seconds in our test, two to three times faster than the GPT Image tier Dimly uses by default.
Range of looks
Film photography, painterly light, paper-cutout sets and neon nights from plain-language prompts, so a whole video can hold one visual language.
Chinese and English text
MiniMax pitches it on reasoned text rendering for signs and posters in both languages. Short words came out right in our tests, though it sometimes adds small made-up text around them.
Open weights
MiniMax publishes the weights: a 33-billion-parameter diffusion transformer that draws in six steps, from the same family as the H3 video model.
Technical specifications
H3 Image
MiniMax's open-weights text-to-image and edit model
16 references
Public image URLs, up to 20 MB each
1K / 2K / 4K
About 1, 4 and 16 megapixels, all native
21:9 to 9:16
Eleven aspect ratios, or a custom size
~23 s
A 2K frame with two references, measured
PNG or JPEG
One image per request
5,000 characters
Prompt length, Chinese or English
6 steps
No negative prompt; describe what you want in the frame
Prompt examples
Night bus
She sits alone at the back of an empty night bus, bass case on the seat beside her, orange sodium streetlights streaking across the windows, looking out at the city
Edit promptRooftop at dawn
She stands on a tar rooftop at dawn playing the bass, the city skyline soft and pink behind her, wind in her hair. Vertical frame, full figure, natural light
Edit promptDesert blue hour
A lone white tour van parked on an empty desert highway at blue hour, headlights on, a single figure sitting on the roof looking at the first stars
Edit promptNeon sign
A rain-slicked alley at night with a red neon sign over a diner door that reads exactly "OPEN LATE". No other words anywhere in the frame
Edit promptPaper set
A paper-cutout stop-motion set: a cardboard moon hanging on a string above layered paper waves, a tiny paper boat with a glowing lantern
Edit promptForest choir
A dozen people in white shirts singing together in a misty pine forest in the morning, light beams through the trees, shot from low
Edit promptUse cases

Keyframes that keep the cast
Draw each shot of a verse from the same character sheet, and the singer looks like the same person on the bus, on the stage and on the roof.

Live-show frames
Stage light, haze and a crowd in silhouette for the performance shots between the story scenes.

Wide establishing shots
Open a video on a place, like a desert road at blue hour, before the first face appears.

Character sheets
A clean, well-lit portrait of the lead is the reference every later frame is drawn from.
Overview
MiniMax H3 Image is MiniMax's open-weights image model, built from the same family as its H3 video model. It turns a prompt into a picture, or edits from up to sixteen reference images at once, at native 1K, 2K or 4K. It is a 33-billion-parameter diffusion transformer that draws in six steps, which is why a 2K frame takes seconds rather than a minute.
Why sixteen references matter for a music video
A music video is dozens of frames of the same people in different places. Each frame needs the character sheet for every person in the shot, the location, and often the previous frame, so the light carries over. Most image models cap references at eight or fewer. Sixteen leaves room for a two-person shot without dropping anything.
The four bassist frames on this page were drawn from one portrait: the same face, haircut, jacket and bass on a night bus, a rooftop, a laundromat and a stage.
Lettering and prompt rewriting
Kie, the API Dimly calls it through, rewrites prompts into a longer description by default. In our tests that rewriting once dropped a letter from the words we asked for, so Dimly always turns it off. With it off, short words came out right, but the model sometimes adds small made-up text around them, so Dimly keeps keyframes with lyrics in the frame on GPT Image 2.6.
MiniMax H3 Image and Dimly
Every Dimly music video starts as images: character sheets, look previews and one keyframe per shot, which a video model then animates. You don't pick the image model in the studio. Dimly's director draws with GPT Image 2.6 by default and can be switched to MiniMax H3 Image, which costs half the credits per keyframe and draws two to three times faster.
For the motion, see MiniMax H3, the video model from the same family.
Simple pricing
Plan your video for free, then pick a plan that fits your release schedule.
Starter
A steady month of chorus cuts, or one shorter song.
- 600 credits, valid for 30 days
- ≈ 8 chorus cuts (15 s)
- or one song up to 2:20
- Every model and resolution · 3 renders at once
Pro
One full song a month with room to retake, and bigger sizes when a release needs them.
Credits a month
- 1,000 credits, valid for 30 days
- ≈ 14 chorus cuts (15 s)
- or 1 full song (3:45)
- Every model and resolution · 4 renders at once
Creator
Several full songs a month, for a steady release schedule.
- 5,000 credits, valid for 30 days
- ≈ 72 chorus cuts (15 s)
- or ≈ 5 full songs (3:45)
- Every model and resolution · 8 renders at once
Full song, chorus cut, lyric video or visualizer · Watermark-free exports on every plan · Export rights depend on your source permissions and provider terms
Credit packs
Pay once, no subscription. Made for one AI effect or one song; if you will make more, a plan above is the better rate.
Single
One AI effect video, like Hotel Lobby or AI Rap Duo.
- 250 credits, valid for 1 year
- 1 AI effect video (Hotel Lobby, AI Rap Duo)
- or ≈ 3 chorus cuts (15 s)
- Exports carry the Dimly watermark
- No subscription, never renews
Song
One full music video (a song up to 3:45), or four AI effects.
- 1,000 credits, valid for 1 year
- 4 AI effect videos (Hotel Lobby, AI Rap Duo)
- or 1 full song (3:45)
- Watermark-free exports
- No subscription, never renews
Studio
A full song in 2K with room for retakes, or two songs.
- 2,000 credits, valid for 1 year
- 8 AI effect videos (Hotel Lobby, AI Rap Duo)
- or ≈ 2 full songs (3:45)
- Watermark-free exports
- No subscription, never renews
Estimates use MiniMax H3 at 768P with 5-second shots (a 15-second chorus cut counts as three), a keyframe per shot, two reference sheets and no retakes; a full song means 3:45, the typical length of the songs people bring. Higher resolutions and the other models cost more per video. You see the exact quote before anything renders, and failed renders are refunded.
FAQs
MiniMax's open-weights image generation and editing model. It takes a prompt and up to sixteen reference images and draws one image at native 1K, 2K or 4K.
No. MiniMax H3 is the video model; H3 Image is the image model from the same family. In Dimly, an image model draws the keyframes and a video model such as MiniMax H3 animates them.
Not as a choice in the studio: you don't pick the image model. Dimly's director draws character sheets, look previews and keyframes with GPT Image 2.6 by default and can be set to use MiniMax H3 Image instead. Keyframes with lyrics in the frame stay on GPT Image either way.
One to sixteen, passed as public image URLs of up to 20 MB each. MiniMax refers to them in the prompt as Picture 1, Picture 2 and so on, in order.
Native 1K, 2K and 4K (about 1, 4 and 16 megapixels) in eleven aspect ratios from 21:9 to 9:16, or a custom width and height of up to 8,192 pixels a side.
Dimly generated them with MiniMax H3 Image at 2K on 15 November 2026, through the same Kie route the studio uses, and downscaled them for the web. The four bassist frames share one reference image, the rehearsal-room portrait.
Other models
See all modelsMiniMax H3
MiniMax
Reference-driven video with native stereo sound. Your track drives the picture, 768P up to 4K.
Seedance 2.5
ByteDance
First-and-last-frame moves up to thirty seconds, audio-driven at 480p/720p — your song muxed back intact.
Kling 4.0
Kuaishou
Kuaishou's long-take video model: up to thirty seconds in one pass, ten keyframes and Omni Reference with up to fifteen inputs.
Seedream 5.5
ByteDance Seed
ByteDance Seed's image model: natural skin and light, edits that leave the rest of the frame alone, and dense small text that stays legible.




