8-step distilled from MiniMax H3 · audio in the same pass · no install
FastH3 Omni
Describe the shot, add up to 9 reference images and 3 audio references. Get a 5–15s clip at 24fps with synced audio in one pass.
- Reference images
- 9
- Audio references
- 3
- Clip length
- 5.875–14.375s
- Aspect ratios
- 4
Example clip. Not a customer result.
Independent tool. Not affiliated with FastVideo, Nuva Lab, or MiniMax.
01
Generate a clip right here
Every input the model accepts is on this form. Write a prompt, add references if you have them, pick an aspect and length, and watch the clip render.
01Describe the shot
141 / 4000
02Reference images
0 / 9
Pictures 1–3 render in real time. Optional — leave empty for text-to-video. JPEG, PNG or WebP up to 25 MB, aspect between 1:4 and 4:1.
03Audio references
0 / 3
Audio references need at least one reference image.
04First and last frame
Frame control needs at least one reference image.
05Aspect
1344 × 768
06Length
24 fps · audio in the same pass
›Advanced
Video references and clip continuation are not supported by this generator.
59 credits
10 credits per second (12 at 1024×768). Failed runs are refunded.
Output
1344×768 · 5.875s · 24fps
5.875s
Your clip plays here as it renders, then saves as an MP4.
- Connect
- Upload
- Generate
- Save
- Done
Keep this tab open until the MP4 is saved. Clips are kept for 7 days.
02
One model, every reference input
FastH3 Omni conditions one generation on text, images and audio together. This is what the generator accepts, and what it does not.
| Input | Limit | How it is used |
|---|---|---|
| Text prompt | Required | Action, camera and sound. Refer to images as <Picture 1> and to audio as Audio 1. |
| Reference images | 0–9 | Characters, products or places to keep on-model. Pictures 1–3 render in real time; 4–9 take longer. |
| Audio references | 0–3 · 2–15s each | A voice, a sound or a music bed. Needs at least one reference image. |
| First / last frame | Optional | Pin the opening or closing composition. Needs at least one reference image. |
| Video reference | Not supported | The hosted model does not take a video as input. Use a still from it as a reference image. |
| Clip continuation | Not supported | Extending a previous clip is not offered in this generator yet. |
03
FastH3 vs FastH3 Omni
Same family, different jobs. FastH3 was the first fast preview; Omni is the version that takes references.
| Compared | FastH3 | FastH3 Omni |
|---|---|---|
| Inference steps | FastH34-step preview | Omni8 steps |
| Distilled from | FastH3MiniMax H3, text-to-video with audio | OmniMiniMax H3 Ref2VA (reference-to-video with audio) |
| Reference images | FastH3None | OmniUp to 9 |
| Audio references | FastH3None | OmniUp to 3 |
| Audio in the output | FastH3Yes, generated with the picture | OmniYes, generated with the picture |
| Best for | FastH3Quick text-only drafts | OmniKeeping a character, product or voice consistent |
04
What a reference actually controls
Each kind of reference steers a different part of the clip. Name them in the prompt so the model knows which is which.
Images keep things on-model
A face, an outfit, a product or a room. Use several images of the same subject for a steadier likeness, or different subjects to put them in one shot.
<Picture 1> hands <Picture 2> a coffee across the counter.
Audio sets voice and sound
A voice sample, a sound effect or a music bed. Say what each clip is for; the model does not guess.
Audio 1 is her voice. Audio 2 is the background music.
Frames fix the edit points
A first frame locks the opening composition, a last frame where the shot lands. Useful for cutting clips together.
First frame: the empty street. Last frame: the neon sign.
The prompt directs the rest
Action, camera move, timing and mood. Keep it to one shot; a 5–15s clip has room for one clear idea.
Slow push-in. She looks up and smiles as the train arrives.
05
How FastH3 Omni works
A distilled reference-to-video model renders picture and sound together, fast enough to stream as it generates.
01
Distilled to 8 steps
FastH3 Omni is distilled from MiniMax H3 Ref2VA down to 8 denoising steps, so a short clip renders in seconds instead of minutes.
02
References condition one pass
Your prompt, images and audio go into the same generation. Picture and sound come out together, so lips, footsteps and music line up.
03
Watch it, then keep the MP4
The clip plays in your browser while it renders. When it finishes we save an MP4 to your account for 7 days.
06
Output specs
Four frame sizes and five fixed lengths. Lengths are exact model values, not rounded.
Aspect and size
16:9
1344 × 768
1:1
768 × 768
9:16
768 × 1344
4:3
1024 × 768
Length and price
| Length | 768-class sizes | 1024 × 768 |
|---|---|---|
| 5.875s | 59 credits | 71 credits |
| 8s | 80 credits | 96 credits |
| 10.125s | 102 credits | 122 credits |
| 12.25s | 123 credits | 147 credits |
| 14.375s | 144 credits | 173 credits |
- Frame rate
- 24 fps from the model
- Audio
- Generated with the picture, in the same file
- File
- MP4 (H.264 + AAC), kept 7 days
07
Pricing
Credit packs, paid once. 1 credit is $0.01; a clip costs 10 credits per second, or 12 at 1024 × 768.
Starter
$9.90
1,200 credits · about 20 short clips
Enough to learn the model: try references, audio and every aspect.
Buy StarterPro
$19.90
2,800 credits · about 47 short clips
Best value per credit: 2.8× the credits of Starter for 2× the price.
Buy ProStudio
$49.00
7,500 credits · about 127 short clips
For regular production work and long clips at 1024 × 768.
Buy StudioPacks do not expire. A run that fails is refunded automatically. Checkout by Stripe.
08
FastH3 Omni questions
What is FastH3 Omni?
FastH3 Omni is an open-weight video model from FastVideo (Hao AI Lab) and Nuva Lab, distilled from MiniMax H3 Ref2VA to 8 steps. It turns a prompt plus optional reference images and audio into a 5–15 second clip with sound. This site runs it in your browser.
Is this the official FastH3 Omni site?
No. This is an independent tool that runs the model for you. It is not affiliated with FastVideo, Nuva Lab, or MiniMax.
How many references can I use?
Up to 9 reference images and up to 3 audio clips per generation. Audio references need at least one image. Pictures 1–3 render in real time; pictures 4–9 make the run slower.
Can I generate without a reference image?
Yes. Leave the image slots empty and the prompt alone drives the clip, sound included. Audio references and frame control do need an image.
Does it accept video references?
No. The hosted model takes images and audio, not video. Export a still from your video and use it as a reference image instead.
What lengths and sizes can I make?
Five fixed lengths: 5.875, 8, 10.125, 12.25 and 14.375 seconds. Four sizes: 1344 × 768 (16:9), 768 × 768 (1:1), 768 × 1344 (9:16) and 1024 × 768 (4:3). The model renders at 24 fps.
How much does one clip cost?
10 credits per second at the 768-class sizes and 12 credits per second at 1024 × 768. A 5.875-second clip is 59 credits ($0.59); the longest 14.375-second clip at 1024 × 768 is 173 credits ($1.73). The price is shown on the Generate button before you run.
Is there a free trial?
No. Generation needs credits from a pack. You only pay for clips that finish: if a run fails, its credits come back automatically.
How long does a clip take?
With up to 3 reference images, a 5.875-second clip is usually saved within 20 seconds; much of that is the clip playing as it renders. More reference images and longer clips take longer. Keep the tab open until the MP4 is saved.
How long do you keep my clips?
Finished MP4s stay in your account for 7 days, then they are deleted. Download anything you want to keep. We do not train models on your uploads.
09
Make your first clip
Write a prompt, add a reference, and see it render in seconds.
