MiniMax H3 Max Reference
MiniMax H3 Max Reference is the reference-guided mode of MiniMax's H3 Max video model, accepting up to 5 reference images, 3 video clips, and 3 audio clips alongside a prompt, plus optional start and end frames that the clip opens and closes on exactly. A reference video is followed shot for shot, in framing and timing, while the prompt changes the subject or the action in it, so an existing clip or ad can be edited whole with a new subject that moves and acts inside the original shots. Image references keep characters and styles consistent, and output is 480p, 768p, or 1080p clips at 5–15 seconds with native stereo audio and selectable or adaptive aspect ratios from 21:9 to 9:16.
See more
MiniMax H3 Max Reference is the reference-guided mode of MiniMax's H3 Max video model, accepting up to 5 reference images, 3 video clips, and 3 audio clips alongside a prompt, plus optional start and end frames that the clip opens and closes on exactly. A reference video is followed shot for shot, in framing and timing, while the prompt changes the subject or the action in it, so an existing clip or ad can be edited whole with a new subject that moves and acts inside the original shots. Image references keep characters and styles consistent, and output is 480p, 768p, or 1080p clips at 5–15 seconds with native stereo audio and selectable or adaptive aspect ratios from 21:9 to 9:16.
Connect Claude, Cursor, VS Code, or Codex to Layer’s MCP server and MiniMax H3 Max Reference becomes a tool your agent can reach for — it picks the parameters, runs the generation, and shows you the result in the conversation.
https://mcp.app.layer.ai/mcpAsk your agent
“Use MiniMax H3 Max Reference to animate a torch-lit dungeon corridor with a slow dolly forward.”
execute_forge({
"base_model_id": "minimax-h3-max-reference",
"prompt": "A torch-lit dungeon corridor, slow dolly forward",
"batch_size": 1,
"duration_seconds": 5
})Authenticate with a personal access token, post the model’s own form, and poll the inference until it completes. Every field below is accepted by this endpoint.
# Start the generation — returns an inference id straight away.
curl -X POST https://api.app.layer.ai/api/v2/workspaces/$WORKSPACE_ID/base-models/minimax-h3-max-reference/inferences \
-H "Authorization: Bearer $LAYER_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"prompt": "A torch-lit dungeon corridor, slow dolly forward"
}'
# Poll until status is "complete" — the response then carries the asset URLs.
curl https://api.app.layer.ai/api/v2/workspaces/$WORKSPACE_ID/inferences/$INFERENCE_ID \
-H "Authorization: Bearer $LAYER_TOKEN"- prompt stringrequired
- Describe what to generate, or the change to make.
- aspect_ratio enumdefault LANDSCAPE_16_9_480p22 options
- The shape and size of the output. Auto follows the shape of the input.
- prompt_enhancement enumdefault AUTO5 options
- Let the model rewrite the prompt before generating.
- seed integer
- The starting point for the model's randomness. The same seed and settings give the same result.
- duration_seconds integerdefault 5
- Length of the generated clip, in seconds.
- guidance_file_first_frame file[]
- The frame the video starts on
- guidance_file_last_frame file[]
- The frame the video ends on
- guidance_file_reference_image file[]
- Guides style and content without being modified
- guidance_file_reference_video file[]
- Guides style and motion without being modified
- guidance_file_reference_audio file[]
- Guides style and voice without being modified
- prompt_language string
- The language the prompt is written in, so it reaches the model as you meant it.
Why use MiniMax H3 Max Reference for video?
Cinematic video generation
MiniMax H3 Max Reference creates high-fidelity video content suitable for trailers, gameplay previews, and marketing materials with temporal consistency.
Flexible input modes
Generate from text prompts, reference images, or existing video. Supports various aspect ratios and durations for different content needs.
Production-ready output
Clean motion, stable framing, and coherent transitions make outputs ready for direct use in UA campaigns and social content.
Game trailers and cinematic previews
Create attention-grabbing trailers and gameplay previews that showcase your game at its best. High temporal consistency ensures smooth, professional results.
UA video creative at scale
Generate dozens of video ad variants for A/B testing across platforms. Rapid iteration lets performance teams optimize creative angles quickly.
Character animations and in-game cinematics
Bring characters to life with animated sequences for cutscenes, promotional materials, and social content.
FAQ
How much does MiniMax H3 Max Reference cost? Is it free?+
How does Layer's pricing work?+
There are no seat fees, feature gates, or plans on Layer. Instead, our platform uses a consumption based system with Creative Units (CUs) with a flexible monthly subscription. Every generation on Layer (image, video, 3D, or audio) consumes a Creative Unit, and you only pay for what you create.
What Elo score does MiniMax H3 Max Reference have?+
Other video models on Layer
MiniMax H3 Reference
MiniMax
Reference-guided MiniMax H3 video at 2K that follows a reference clip shot for shot while the prompt changes the subject or action.
MiniMax H3
MiniMax
MiniMax H3 (Hailuo-03) next-gen open-weight video with native stereo audio at 2K and 24 FPS.
Minimax Hailuo-2.3 Standard
MiniMax
Latest standard video model. Improved prompt understanding and visual consistency for daily creation.
MiniMax H3 Max Turbo
MiniMax
MiniMax H3 Max Turbo post-trained video with native stereo audio at 480p, 768p, or 1080p.
MiniMax H3 Max Insert
MiniMax
Insert a new 5-13 second scene into a video with MiniMax H3 Max, then return to the original footage.
Seedance 2.5 Reference
ByteDance
Reference-guided Seedance 2.5 video with up to 30 images, 10 videos, and 10 audio files.