LTX 2.5
Lightricks
Explore modelMiniMax / video model
MiniMax H3
Explore MiniMax H3, its capabilities, use cases and ComfyUI workflows for video generation on HexGrid Cloud.
Deploy MiniMax H3 with ComfyUI
Choose a workflow to explore its setup, inputs, downloads and generation examples.
01Workflow guide
Text to Video
Create video from text prompts with MiniMax H3. Explore the ComfyUI setup, workflow instructions and generation examples.
Text → Video
Explore workflow02Workflow guide
Image to Video
Animate an input image into video with MiniMax H3. Explore the ComfyUI setup, workflow instructions and generation examples.
Text + Image → Video
Explore workflow03Workflow guide
Reference to Video
Guide video generation with reference media with MiniMax H3. Explore the ComfyUI setup, workflow instructions and generation examples.
Text + Image + Video + Audio → Video
Explore workflowAbout MiniMax H3
MiniMax H3 is MiniMax's open-weight, multimodal video generation model, deployable on HexGrid.cloud with dedicated GPU infrastructure, configurable inference workflows and support for custom model configurations.
It supports three primary generation modes: text-to-video, image-to-video and reference-to-video. The model accepts text, images, video clips and audio references, generating video with native synchronized stereo audio, including dialogue, environmental sounds and music.
MiniMax H3 uses a 33-billion-parameter Omni Transformer architecture with unified multimodal understanding. It supports 4–15 second clips at 24 FPS, native 768p generation and up to 2K output through an additional regeneration stage. Its multimodal reference system enables character consistency, motion transfer, voice conditioning and instruction-based video editing.
What people use it for
Cinematic text-to-video
Generate cinematic videos directly from detailed text prompts, controlling characters, actions, environments, camera movements, lighting, dialogue and synchronized environmental audio.
Image animation and transitions
Animate existing images into cinematic videos using first-frame, last-frame or two-keyframe conditioning. Control how subjects, environments and camera composition evolve between the supplied frames.
Consistent characters and references
Generate videos using multiple reference images, video clips and audio samples to maintain character identity, visual style, motion characteristics and voice consistency across scenes.
Motion transfer and video editing
Use reference videos to guide character movements, camera behavior and scene dynamics. Transform existing footage through natural-language instructions while preserving selected subjects, actions or visual characteristics.
Advertising and product videos
Create promotional videos, product showcases, branded animations and cinematic marketing content with detailed visual compositions, readable text, consistent product appearance and synchronized sound.
Strengths
- Native 768p generation with up to 2K output through context-aware video regeneration
- 4–15 second video generation at 24 FPS with native 32 kHz stereo audio
- Three primary generation modes: text-to-video, image-to-video and reference-to-video
- Unified multimodal understanding across text, images, video and audio references
- Up to 9 reference images, 3 reference videos and 3 reference audio clips in a single generation request
- First-frame, last-frame and first-to-last-frame conditioning for controlled cinematic transitions
- Consistent subject identity, clothing, visual style and environmental characteristics using reference inputs
- Motion transfer from reference videos, including character actions, camera movements and scene dynamics
- Native synchronized dialogue, sound effects, environmental audio and music generation
- Voice conditioning from reference audio for character speech and dialogue generation
- Native multi-shot generation with continuity across connected cinematic sequences
- Strong instruction following for complex multimodal prompts and reference relationships
- Improved text rendering, product branding and typography within generated videos
- Contextual Omni Representation for understanding relationships between reference materials and target scenes
- H3-Regenerate-2K for context-aware high-resolution refinement of generated videos
- Open-weight architecture supporting custom inference pipelines, model modification and fine-tuning
Limitations
- The publicly released H3-Base checkpoints generate at 768p. Producing 2K video requires the separate H3-Regenerate-2K stage, which is not included in the open-weight release.
- 4K generation is not supported natively by the official H3 model; higher-resolution output requires external upscaling or post-processing.
- Individual video generations are limited to 4–15 seconds. Longer videos require multiple generations, video editing or additional stitching workflows.
- Output frame rate is fixed at 24 FPS in standard generation. Higher playback frame rates require external frame interpolation.
- Reference-to-video supports a maximum of 9 images, 3 video clips and 3 audio clips, with no more than 12 reference files combined.
- Reference video and audio clips must each be 2–15 seconds long, with a maximum combined duration of 15 seconds for each media category.
- Audio references cannot be used as the only input in reference-to-video mode; at least one image or video reference is required.
- First-and-last-frame generation accepts a maximum of two conditioning images, while larger multimodal reference sets require the separate reference-to-video checkpoint.
- H3-Context-IR, the official multimodal instruction-processing component, is not included in the open-weight release and requires a separate integration or alternative preprocessing workflow.
- Context-aware 2K regeneration introduces additional processing overhead compared with direct 768p generation.
- Complex multi-character interactions, precise physical movements and extended scene continuity may exhibit temporal inconsistencies or visual artifacts.
- The 33B-parameter architecture requires substantial GPU memory and computational resources, with multi-GPU execution recommended for full-precision inference.
- The MiniMax H3 Community License imposes territorial and commercial-use conditions. Its standard grant excludes the United States, European Union, United Kingdom and South Korea, and certain commercial uses require separate authorization.
Specifications
Generation, input and serving details.
ProviderMiniMaxFamilyH3CategoryvideoWorkflows3
Generation
- Modes
- Text-to-video, image-to-video, reference-to-video
- Variants
- H3-Base-FL2VA, H3-Base-Ref2VA
- Architecture
- H3-Omni Transformer with unified multimodal audio-video generation
- Model parameters
- 33B-parameter dense Omni Transformer
- Text encoder
- H3-Encoder based on Qwen3-VL-32B
- Resolution
- 768p base generation; up to 2K through H3-Regenerate-2K
- Default resolution
- 768p for open-weight base inference
- Duration
- 4–15 seconds
- Duration options
- 4–15 seconds, configurable in whole-second increments
- Frame rates
- 24 FPS
- Default frame rate
- 24 FPS
- Aspect ratios
- 21:9, 16:9, 4:3, 1:1, 3:4, 9:16; adaptive framing for reference-conditioned workflows
- Audio generation
- Native synchronized stereo audio, including dialogue, sound effects, environmental audio and music
- Audio sample rate
- 32 kHz stereo
- Dialogue languages
- 11 reliably supported languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish
- Multi-shot
- Native multi-shot video generation with contextual continuity
- Video decoder
- H3-VisualVAE with temporal and spatial compression
- Audio decoder
- H3-AudioVAE with separate stereo-channel processing
- High-resolution rendering
- H3-Regenerate-2K using context-aware in-context regeneration
Inputs and Controls
- Text prompt
- Required for text-to-video; used to direct image-to-video and reference-to-video generation
- Prompt length
- Up to 7,000 characters
- Start image
- Optional for first-frame-conditioned image-to-video generation
- End image
- Optional for last-frame-conditioned image-to-video generation
- Keyframe conditioning
- First frame, last frame or both; maximum 2 conditioning images
- Reference images
- Up to 9 images
- Reference videos
- Up to 3 clips, each 2–15 seconds; maximum 15 seconds combined
- Reference audio
- Up to 3 clips, each 2–15 seconds; maximum 15 seconds combined
- Total reference files
- Maximum 12 files across images, videos and audio
- Audio-only references
- Not supported; reference audio must be accompanied by at least one image or video
- Image dimensions
- 256–5760 pixels per dimension for supported reference inputs
- Image formats
- JPG, JPEG, PNG, WEBP, HEIC, HEIF
- Video formats
- MP4 or MOV using compatible H.264 or H.265 encoding
- Audio formats
- WAV, MP3
- Camera motion
- Natural-language control of pan, tilt, zoom, dolly, tracking, orbital movement and focus transitions
- Motion reference
- Supported through reference video conditioning
- Character consistency
- Reference-guided subject identity, appearance, clothing and visual style preservation
- Voice conditioning
- Reference audio can guide speaker timbre, dialogue characteristics and vocal style
- Video editing
- Instruction-based editing using existing video and multimodal references
- Prompt expansion
- Supported through H3-Context-IR or compatible custom preprocessing workflows
- Multimodal prompting
- Natural-language descriptions of relationships between text, images, videos and audio references
- Inference steps
- Configurable through compatible inference workflows
- Guidance
- CFG-distilled base checkpoints with workflow-dependent guidance controls
- Random seed
- Configurable integer seed for reproducible generation
- Frame interpolation
- Optional through compatible post-processing workflows
- Reference indexing
- Reference materials can be identified and described by their position or role within multimodal prompts
- Model customization
- Weight modification and fine-tuning supported; implementation depends on the training and inference workflow