Orbit left
The camera arcs around the subject, holding it centred while the background slides to the right.
Best forRevealing a product's side profile or the space behind a subject.
64Credits
Remaining credits：
Public
Fast H3 Max generation · four creation modes on Kuca
MiniMax H3 Max: faster generation, precise camera control
MiniMax H3 Max is tuned for creators who want polished video with less waiting. It follows detailed prompts closely, produces visually refined frames, and supports text, image, mixed-reference and camera-control workflows. On Kuca, you can move from brief to preview faster while keeping 480P–1080P output, 5–15 second clips and six ready camera moves in one workspace.
- Speed
- High-throughput generation
- References
- 9 images · 3 videos · 3 audio
- Camera
- 6 preset trajectories
- Output
- 480P–1080P · 5–15s
H3 or H3 Max? Read the differences before you spend credits
Both models belong to the same H3 generation family, so the choice comes down to what each one was tuned for. H3 Max leans toward prompt adherence, aesthetics and camera precision. MiniMax H3 remains the 2K model with jointly generated stereo sound.
| Dimension | MiniMax H3 | MiniMax H3 Max |
|---|---|---|
| Tuned for | 2K detail and native stereo audio | Prompt adherence, visual aesthetics and camera precision |
| Resolution | 768P, 2K | 480P, 768P, 1080P |
| Duration | 4–15 seconds | 5–15 seconds |
| Reference material | Up to 5 images and 3 audio clips | Up to 9 images, 3 videos and 3 audio clips, 12 files in total |
| Camera control | Written into the prompt | Six selectable trajectory presets |
| Generated audio | Native stereo, generated with the picture | Not generated — audio enters as a reference |
| Prompt length | Up to 7,000 characters | Up to 50,000 characters |
| Throughput | Baseline H3 pipeline | High-throughput pipeline for faster generation |
Open the MiniMax H3 page
Both pages open the same generator, so you can move between H3 and H3 Max without rebuilding a brief.
Six camera moves you select instead of describe
Choose one of six ready camera moves for the shot: orbit left, orbit right, dolly in, dolly out, crane up or crane down. Each preset turns camera direction into a clear setting, so your prompt can stay focused on the subject, action and scene.
Orbit right
The same arc in the opposite direction, with the background sliding to the left.
Best forMatching the direction of existing footage or following on-screen movement.
Dolly in
The camera pushes straight toward the subject and the frame tightens around it.
Best forBuilding emphasis on a face, a label or a piece of detail.
Dolly out
The camera pulls back and the surroundings enter the frame.
Best forClosing a beat by revealing where the subject actually is.
Crane up
The camera rises and the view tilts down onto the subject.
Best forOpening a scene or showing the scale of a location.
Crane down
The camera descends toward the subject's own level.
Best forSettling into an intimate, eye-level finish.
Camera controls take one first frame and one preset, and the prompt is optional in that mode. Aspect ratio follows the frame you upload.
Four endpoints, four different contracts
Each H3 Max endpoint asks for different inputs and leaves different decisions open. Choosing the endpoint first is what keeps a reference set small.
- 01
Text to video
- You supply
- A prompt and nothing else.
- The model decides
- Framing, subject, lighting and camera.
- Use it when
- You are still exploring a direction and nothing has to be preserved.
- 02
Image to video
- You supply
- A first frame, plus an optional ending frame.
- The model decides
- Motion and timing, while framing follows your image.
- Use it when
- Composition, identity or product geometry has to survive the render.
- 03
Mixed reference
- You supply
- Images, video clips and audio clips, each with a named role.
- The model decides
- How those references combine into a single shot.
- Use it when
- Motion, style or sound has to come from existing material rather than a description.
- 04
Camera controls
- You supply
- A first frame and one of the six presets.
- The model decides
- Subject behaviour inside the move, because the trajectory comes from the preset.
- Use it when
- The shot exists mainly to show a specific camera movement.
Every reference file should own exactly one attribute
H3 Max accepts a lot of material, but a small set with clear ownership beats a large set that argues with itself. Give each file one job and name that job in the prompt.
| Asset | What it should own | Limit | Common mistake |
|---|---|---|---|
| Image | Identity, product shape, wardrobe, environment or palette. | Up to 9 images. | Uploading several near-identical frames that compete for the same attribute. |
| Video | A camera move, a rhythm, a gesture or the way a material behaves. | Up to 3 clips, 2–15 seconds each, 15 seconds combined. | Feeding in footage with cuts, text overlays or a different subject. |
| Audio | A voice, a music bed, an ambience or a sound effect. | Up to 3 clips, 2–15 seconds each, 15 seconds combined. | Expecting the clip to be regenerated: audio is guidance, not output. |
| Prompt | Subject, place, action, camera, chronological order and boundaries. | Up to 50,000 characters. | One vague sentence, or three separate plots squeezed into five seconds. |
Images, videos and audio share a combined ceiling of 12 files, and reference mode needs at least one of them.
The exact envelope every request is validated against
These values come from the generator's own configuration. It is quicker to check them while you assemble a shot than after a rejected submission.
- Duration
- 5–15 seconds, whole seconds only
- Resolution
- 480P, 768P or 1080P, defaulting to 768P
- Aspect ratio, text to video
- 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
- Aspect ratio, reference mode
- The six ratios above, plus adaptive
- Aspect ratio, image and camera modes
- Follows the first frame you upload
- Image references
- Up to 9
- Video references
- Up to 3 clips, 2–15 seconds each, 15 seconds combined
- Audio references
- Up to 3 clips, 2–15 seconds each, 15 seconds combined
- Reference assets in total
- 12 maximum, and at least 1 in reference mode
- Prompt length
- Up to 50,000 characters
Camera controls accept a first frame and a preset, and the prompt is optional for that endpoint.
Five failures that waste the most takes
Most disappointing H3 Max renders trace back to the brief or the reference set rather than to the model. These are the patterns worth checking first.
The render ignores what the prompt asked for
Why it happensThe brief is either too vague or contains several unrelated ideas.
What to changeDescribe one action with concrete nouns, then state the camera behaviour in a separate sentence.
The shot feels rushed or cuts off mid-motion
Why it happensToo much plot is packed into a 5–15 second window.
What to changeKeep one main action, drop the secondary beat, or extend the duration before adding detail.
Text, logos or signage come out warped
Why it happensA source image already contains lettering, watermarks or busy background clutter.
What to changeStart from a clean frame, crop the lettering out, or keep on-screen text out of the shot entirely.
The subject drifts away from the reference
Why it happensTwo files are quietly competing for the same attribute, such as identity or colour.
What to changeGive each attribute a single owner and say so in the prompt, then remove the redundant file.
The camera ignores the movement you described
Why it happensA precise move was written as prose in an endpoint that expects a preset.
What to changeSwitch to camera controls, choose one of the six trajectories, and let the prompt cover subject behaviour only.
Four shots H3 Max handles well
Each playbook lists what to prepare, how to shape the prompt and what usually goes wrong. Load any prompt straight into the generator and adapt it to your own footage.
Vertical social cut
- Prepare
- One clean 9:16 or 3:4 subject image, plus a short reference clip when a specific move matters.
- Prompt shape
- Subject, then place, then a single action, then the camera, then the beat you want in the final second.
- Watch out
- Cramming a full story into five seconds. Keep one action and let it finish.
Product turn
- Prepare
- A well-lit product image on a plain surface, ideally without printed text or reflective clutter.
- Prompt shape
- Product, then surface, then the camera preset, then notes on lighting and material.
- Watch out
- Lettering on packaging is the first thing to warp. Crop it out or accept a softer logo.
Open-world action beat
- Prepare
- Character and environment images, plus a short clip that demonstrates the movement you want.
- Prompt shape
- Character, then environment, then one action, then the camera, then continuity boundaries.
- Watch out
- Reusing recognisable game or film assets invites IP trouble. Keep every design original.
Trailer-style teaser
- Prepare
- A hero image and a music reference, rendered as several short takes instead of one long shot.
- Prompt shape
- A shot list of 5–15 second beats, each with its own camera preset and a single action.
- Watch out
- One request cannot carry a whole trailer. Build it take by take and cut it together in an editor.
Keep source material original or properly licensed. Do not upload game assets, film frames, logos or other protected characters and brands, and avoid prompts that imitate a specific franchise.
Questions people ask before their first H3 Max render
MiniMax H3 Max is a faster, more controllable version of the H3 video model. It improves prompt adherence and visual polish while adding mixed image, video and audio references plus six ready camera moves.
Two qualities improved: prompt adherence, meaning how faithfully the render follows a written brief, and visual aesthetics, meaning how the finished frame is lit, composed and graded. H3 Max is not a larger model trained purely for speed.
H3 Max is built for higher-throughput inference, so it can process a shot faster than the baseline H3 workflow. Actual completion time still varies with resolution, duration, reference files and queue load.
H3 Max responds to prompts more closely, accepts video references alongside images and audio, and offers six preset camera trajectories. MiniMax H3 stays the better choice for 2K output and for native stereo audio generated together with the picture.
480P, 768P and 1080P, with 768P as the default, and any whole-second duration from 5 to 15 seconds. H3 Max does not offer the 2K option available on MiniMax H3.
Up to 9 images, 3 video clips and 3 audio clips, capped at 12 files in total. Video and audio clips each run 2–15 seconds, with a 15-second combined budget per category, and reference mode requires at least one file.
Kuca provides six selectable moves: orbit left, orbit right, dolly in, dolly out, crane up and crane down. Camera control takes a first frame plus one preset, and the prompt becomes optional.
No. In Kuca, audio is an input: up to 3 reference clips of 2–15 seconds that can pin a voice, a music bed, an ambience or a sound effect. If you need a soundtrack generated with the picture, use MiniMax H3.
Work through six parts in order: subject, place, action, camera, chronological order and boundaries such as what must not appear. Then keep one main action per 5–15 second shot and change a single variable between takes.
Vague prompts, several plots packed into a short duration, source images that already contain text or distracting clutter, and reference files that compete for the same attribute. The diagnostics section on this page covers each one and what to change.
Start an H3 Max shot
Choose a mode, add your brief or references, and start a faster H3 Max generation directly on Kuca.