FL2VA
H3-Base-FL2VA takes text alone, one image as the first or last frame, or two images as first and last. Output is video plus audio. MiniMax also calls the text-only path T2VA.
MiniMax H3 · Meta ads
Use MiniMax H3 with a structure card for Meta ads
MiniMax H3 is MiniMax's open multimodal video model. SameHook sits in front of the prompt: turn a winning ad into a hook, beats, and CTA you can edit, then decide what a video model is allowed to make.
SameHook does not ship MiniMax weights, and this studio does not call H3. No ROAS promise. Source frames, music, likeness, and trademarks stay out.
What H3 is
An open video model. Not a SameHook feature.
MiniMax open-sourced H3 on 3 August 2026. It reads text, images, video, and audio together, and it writes video with native stereo audio. MiniMax describes it as a general-purpose system. SameHook does not host it and does not replace it.
- Duration
- 4–15 seconds
- Frame rate
- 24 fps
- Audio
- 32 kHz stereo, generated with the picture
- Resolution
- Open base weights target a 768px shorter side. A separate H3-Regenerate-2K step can reach 2K. That 2K module is not in the open-source weights.
- Aspect
- Includes 9:16, 1:1, 3:4, 4:3, 16:9, and 21:9. MiniMax says the list is not closed.
- Dialogue
- Stable support for 11 languages, including English and Chinese.
Two open checkpoints
Ref2VA
H3-Base-Ref2VA takes references: up to 9 images, 3 video clips, and 3 audio clips. Audio cannot be the only input. Each clip is 2–15 seconds, total reference time stays within 15 seconds, and the file count stops at 12.
Two pieces stay off the open weights. H3-Context-IR is MiniMax's hosted preprocessor for messy multimodal instructions. H3-Regenerate-2K is the in-context 2K pass. MiniMax documents both through their API, not as files in the Hugging Face repo.
MiniMax's own pages
Before the prompt
A structure card is the brief H3 does not write for you.
H3 follows instructions. It does not know which three seconds should stop a Meta scroll, which beat carries the offer, or which CTA you are willing to test. A prompt that says “make it like this ad” is a request to copy. The card cuts those two jobs apart.
01Hook
The stop. One line or one action, in the first one to three seconds. That is the only thing you are trying to keep.
02Beats
The job of each turn. Not a shot list of someone else's frames.
03CTA and voiceover
The ask, and the order of the offer, in your words. Spoken lines, not a cloned voice.
04Aspect and length
Pick 9:16, or another ratio you will actually test, and a duration inside H3's 4–15 second window before you generate.
05Variants
Persona, angle, and CTA are separate knobs. The next render should be a new creative, not a new skin on the same ad.
SameHook
Reference, structure, edit, variants.
The studio already runs this loop. A video model such as H3 is a later render path. Today the download is a preview, not an H3 file.
- 1
Reference
Curated, upload, or link
Start from a structure we wrote, a file on your machine, or a page link. Uploads stay in the browser. The link reader keeps a title and a description, not the page.
- 2
Structure card
Hook, beats, CTA
Hook, beats, voiceover skeleton, CTA, landing intent, aspect, and length. Every field can be rewritten.
- 3
Edit
Offer, person, product
Swap what should be yours. The opening that earned the stop can stay. The original picture does not.
- 4
Variants
Preview, not an H3 file
At least three preview variants: a storyboard, a script, and a placeholder frame. If the server has an AI Gateway key, a text model can draft the spoken lines. The frames stay placeholders either way. Nothing in this build sends the card to MiniMax H3.
FAQ
MiniMax H3, asked directly.
What is MiniMax H3?
MiniMax's open general-purpose video model. Text, images, video, and audio can go in. Video with native stereo audio comes out, about 4–15 seconds at 24 fps. The open checkpoints are FL2VA and Ref2VA, published as MiniMaxAI/MiniMax-H3. The hosted context preprocessor and the 2K regenerate step are separate from those weights.
How is a structure remake different from generating with H3?
H3 makes pixels and sound from a prompt and from references. A structure remake keeps the job of the hook and drops the picture. SameHook is that second job. H3 can be a way to render later. It is not a license to clone the reference.
Can SameHook generate video with MiniMax H3?
Not in this build. The studio preview-renders storyboards, scripts, and placeholder frames. With an AI Gateway key, a text model can draft spoken lines. Frames stay placeholders. No MiniMax provider is wired in.
What are FL2VA and Ref2VA?
FL2VA is first-and-last-frame mode: text, one image, or two images, then video and audio. Ref2VA is reference mode: images, short video clips, and audio that rides along with a picture or a clip. Both are open checkpoints. Neither runs inside SameHook.
Does SameHook replace MiniMax or Hailuo?
No. We do not host H3, we are not Hailuo, and we are not a cheaper H3. For H3 output, use MiniMax's own app, their API, or a runtime you deploy from the open weights.
What should I lock before prompting H3 for a Meta ad?
The hook, the job of each beat, the CTA, the aspect, and a length between 4 and 15 seconds. Write the prompt from that card. Change the persona or the offer when you want another variant. Do not paste the reference ad and ask for the same video.
Do you guarantee the ad will perform?
No. There is no ROAS claim on this page, and there is no performance leaderboard behind it.
Write the card. Then decide how to render it.
Open the studio, edit a structure card, and download the preview variants. If a video model joins the path later, the card is still the brief.
Open the studio