MiniMax H3 Max
Reference to Video
Reference generation is not on the MiniMax H3 Max API yet — MiniMax lists it as coming soon.
It is on MiniMax H3, the base model, which also gives you 2K. That is the engine this page runs, at the same rate per second.
Upload the face, the product or the voice that has to stay the same, then name it in the sentence. MiniMax H3 Max reference to video hands back a 4-to-15-second take that still looks like yours — 768P or 2K, sound in the same file.
Press Try this to load one
5s · 16:9 · 768PA shot brief, beat by beat
Does MiniMax H3 Max Do Reference Yet?
Not yet. MiniMax's own documentation lists reference generation for MiniMax H3 Max as coming soon — today the model does text to video and image to video only. Checked 2 September 2026.
The base model does it, and does it well. MiniMax H3 takes up to nine reference images, three video clips and three audio clips — twelve files in total — and holds their features across the shot. It also outputs 2K, which MiniMax H3 Max cannot.
That is why MiniMax H3 Max reference to video requests land on MiniMax H3 here. The rate does not change: 40 credits a second at 768P, the same as every other page on this site.
Three ways out of here:
· Want the faster engine instead? Text to video runs on MiniMax H3 Max.
· Only anchoring the opening frame? Image to video.
· Want the reference mechanics field by field? That guide lives on our MiniMax H3 site.
- Images · clips · audio
- 9 + 3 + 3
- Files in one request
- 12
- 768P or 2K
- 4–15 s
- Credits a second at 768P
- 40
Images · clips · audio
Files in one request
768P or 2K
Credits a second at 768P
- No 2K on MiniMax H3 Max
- No reference on the Max API
- No surcharge for switching
Name Your References in the Prompt
Uploading is not enough. MiniMax H3 Max reference to video only defends the files you name, so you have to say which one does what. Here is MiniMax's own example, usable as written:
The character performs street dance, following the motion in reference video 1; the character's appearance follows reference images 1 and 2.
reference image 1, 2
Where the look comes from
reference video 1
Where the movement comes from
plain English
No brackets, no @, no field names
Numbers come from upload order
Reorder your files and you have made a different request. Delete one and everything after it shifts up while your prompt still says the old number.
One file can do two jobs
The same portrait can set both the face and the wardrobe. Say both — the model does not assume the second one.
Unnamed files still get sent
They are still uploaded and still counted, and the model may ignore them anyway. If it matters, name it.
Need to control exactly what survives?
MiniMax publishes a second, stricter way to write these prompts: six named fields in a fixed order, four label types, and a set of retention markers that state, per label, whether it is kept, partly kept, transferred onto someone else, or only loosely echoed.
You will not need it for most shots. You will need it when one thing has to survive exactly and everything else has to move.
It is a property of the base model rather than of Max, so the full field-by-field guide is on our MiniMax H3 site.
- subject_definitions
- Names what each file contains
- summary
- Declares the task type
- retention_analysis
- Says what survives, per label
- detailed_description
- The shot, in playback order
- overall_soundscape
- What the room sounds like
- non_diegetic_music
- The score
Rather not phrase it yourself? The prompt generator builds the sentence around whatever you uploaded.
Reference on MiniMax H3 Max vs H3
This is the one tool page here that does not run MiniMax H3 Max. The table is the whole reason.
| MiniMax H3 Max | MiniMax H3 (this page) | |
|---|---|---|
| Reference input | Not yet — listed as coming soon | 9 images · 3 clips · 3 voices |
| Resolution | 480P / 768P | 768P / 2K |
| Duration | 5–15 s | 4–15 s |
| Speed | The fastest thing here | Slower |
| Rate at 768P | 40 credits a second | 40 credits a second — identical |
Switching engines does not cost you anything per second. What you trade for reference support is speed, not rate.
No reference material to work from? Text to video runs on MiniMax H3 Max and comes back faster.
What You Can Upload, and the 12-File Wall
Every number below is from MiniMax's own input requirements table, checked 2 September 2026.
- Images
- ≤ 9 · 30 MB each · 256–5760 px per side · JPG JPEG PNG WEBP HEIC HEIF
- MiniMax input requirements
- Video clips
- ≤ 3 · 50 MB each · 2–15 s each · 15 s combined · H.264 or H.265
- MiniMax input requirements
File sizes are counted per file, not added up — but the whole request has to fit in 64 MB. When your set is bigger than that we switch to URL upload automatically, and nothing about your side of it changes.
What One Reference Clip Costs
A MiniMax H3 Max reference to video job is priced on the seconds you generate, the resolution you picked, and any reference footage you brought.
Five photographs are free. The sixth is not, and a reference clip never was — a clip is read second by second, and billed that way.
| 768P | 2K | |
|---|---|---|
| Per run, flat | None | None |
| Output, per second | 30 credits | 65 credits |
| Reference video, per second | 30 credits | 65 credits |
| Reference images 6–9, each | 40 credits | 40 credits |
| Reference audio | Free | Free |
| 5s out, photos only | 150 credits | 325 credits |
| 15s out, photos only | 450 credits | 975 credits |
| 5s out + a 10s reference clip | 450 credits | 975 credits |
Credits are what you are actually charged. On Pro, 150 credits is exactly $1.00.
- Trim the reference clip before you upload it. Ten seconds of reference costs exactly what ten seconds of output costs. If only three seconds of it carry the movement you want, upload those three — it is the single biggest saving on this page.
- Five photographs are free; the sixth is not. The first five cost nothing at any size, and images six to nine are 40 credits each — fal's $0.08 a still, divided by the same $0.002 every other rate here comes from. Nine photographs instead of five adds 160 credits, more than the five-second 768P render they are attached to. Reference audio adds nothing at all.
- Resolution is the only multiplier. Not the ratio, not the audio, not the batch. Extra photographs and reference seconds are added to the bill, never multiplied into it.
- A run that fails costs nothing, on every plan.
There is no free clip on this page. The free engine cannot take references at all, and the MiniMax H3 Max trial clips run on a model that does not do reference generation yet.
So instead of a token free run, everything before the render is open: upload your files, read the exact price, run the preflight.
Paid starts at $9.9 a month, and commercial use is permitted under the upstream provider's terms.
Same monthly credits either way
- Free
No card, no timer. One clip right now with no account, then three MiniMax H3 Max clips when you sign up.
$0No card, ever
Generate my first clipNo account for the first one
1 clip now, 3 on sign-up · 4 clips total
5s · 480P · with sound after sign-up
Granted on day 0, day 2 and day 7
- Native audio on the three signed-in clips
- No watermark
- Six aspect ratios
- First and last frame
- Saved history
- Commercial use — paid plans only
What's free here and elsewhere → the full breakdown
- Lite
1–6 finished clips a month
$9.90/moBilled yearly, $118.80 today
7-day refund while credits are untouched
1,200 credits a month · 12 clips first try
5s · 768P · 100 credits each
4 at three takes
Same credits yearly or monthly
- Both engines — MiniMax H3 Max and MiniMax H3
- Up to 15 seconds, up to 2K on MiniMax H315s
- No watermark
- Commercial use — permitted under the upstream provider's terms
- Prompt library and prompt help
- One clip at a time, standard queue
- 30-day history · email support
Most expensive per clip here: 83¢ a generation, $2.48 a finished clip. From 7 clips a month, Pro costs less.
- Most popularPro
7–21 finished clips a month
$19.90/moBilled yearly, $238.80 today
7-day refund while credits are untouched
3,000 credits a month · 30 clips first try
5s · 768P · about 66¢ a generation
10 at three takes
Same credits yearly or monthly
- Everything in Lite, plus:
- 4 takes per prompt, one click
- 3 jobs running at once
- Priority queue
- Unlimited history, searchable
- Seed lock + one-word re-roll
- Saved presets and brand kit
- 20% less per credit than Lite
- Priority email support
Past 21 clips a month, Studio costs less.
- Studio
22+ finished clips a month
$59.90/moBilled yearly, $718.80 today
7-day refund while credits are untouched
11,000 credits a month · 110 clips first try
5s · 768P · 54¢ a generation, the lowest here
37 at three takes
Unused credits roll over one month
- Everything in Pro, plus:
- 8 takes per prompt · 8 jobs at once
- Front of the queue
- Credits roll over one month
- 3 seats, one library, one bill
- Project folders · bulk ZIP
- 4-hour support, business days
- New engines first
- Invoices, VAT ID, purchase orders
- API access — private beta waitlist
Under 22 clips? Pro plus a pack is cheaper. We'd rather say so.
Faces, Voices and Characters
Reference material is where the legal risk on this site actually lives. The rules are short. Please read them.
- FacesVideo or photos of a real person need that person's consent. Not public figures, not a stranger's picture.
- VoicesA voice reference clones how one specific person sounds. Use your own, a licensed one, or a synthetic one.
- CharactersA copyrighted character does not become yours because a model regenerated it.
- MinorsNever upload references depicting minors.
- ModerationSubmitted media is screened automatically. False positives and false negatives both happen. A rejected job costs you nothing.
Five Checks Before You Generate
All five run live in the panel beside the Generate button, every time you set up a MiniMax H3 Max reference to video job. You do not have to remember any of them.
Order is part of the request
Files are numbered as you dropped them. Reorder or delete one and the numbering shifts — while your prompt still says the old number.
Order: image 1 · 2 · video 1
Name every file you upload
An unnamed file is still sent, still counted and may still be ignored. You paid to carry it and did not get what you brought it for.
reference video 1 is never named
`adaptive` follows reference image 1
The default frame shape comes from your first image. Pick a ratio yourself if you need a specific one.
Ratio follows reference image 1
Reference clips are billed by their length
Photographs are almost free, but a reference video is read second by second and charged at the output rate. Trim it to the part that carries what you want.
+10s reference = +400 credits
2K only exists on this engine
This is the one page here that outputs 2K, because MiniMax H3 Max stops at 768P. If a shot has to be sharp enough to crop into, make it here.
65 credits a second at 2K
Check 05 is not a warning. It is the thing people on this page most often do not realise they are allowed to do.
How to Use Reference to Video Here
Three steps. The sound arrives inside the same file as the picture.
Upload what must stay the same
Up to nine images, three clips and three voices — twelve files in all. A face, a product, a style, a voice.
≤ 12 filesName them in the sentence
"The character's appearance follows reference images 1 and 2." Naming is what makes the model defend it.
reference image 1Pick 768P or 2K, then download
Four to fifteen seconds, sound already in the file. A run that fails costs no credits.
24 fps · up to 2K
That is the whole browser path: nothing to install, no GPU, no API key.
MiniMax H3 Max Reference to Video FAQ
What runs it, what it costs, and what happens to the files you bring
Does MiniMax H3 Max support reference to video?
Why does this page run on MiniMax H3 instead of H3 Max?
How do I name my reference files in the prompt?
Do reference files cost extra?
Why did my character still drift?
How many files can I upload?
What happens to the files I upload?
Can I get 2K here?
Do I need an account, and is there a free trial?
Can I use the result commercially?
Bring the Face. Keep the Face.
Nine images, three clips, three voices. Twelve files, one sentence, and a take that still looks like yours — 4 to 15 seconds, and up to 2K because it runs on MiniMax H3.
Generate free · queueWritten and maintained by the MiniMax H3 Max AI Video Generator editorial teamPublished Last updated