MiniMax H3
MiniMax
Explore modelQwen / video model
Wan 2.2 14B
Explore Wan 2.2 14B, its capabilities, use cases and ComfyUI workflows for video generation on HexGrid Cloud.
Deploy Wan 2.2 14B 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 Wan 2.2 14B. Explore the ComfyUI setup, workflow instructions and generation examples.
Text → Video
Explore workflow02Workflow guide
Image to Video
Animate an input image into video with Wan 2.2 14B. Explore the ComfyUI setup, workflow instructions and generation examples.
Text + Image → Video
Explore workflow03Workflow guide
First and Last Frame to Video
Create video between a starting image and an ending image with Wan 2.2 14B. Explore the ComfyUI setup, workflow instructions and generation examples.
Text + Image → Video
Explore workflowAbout Wan 2.2 14B
Wan 2.2 A14B is Alibaba's open-weight video generation model, available on HexGrid.cloud through private GPU deployments with preconfigured ComfyUI workflows, custom checkpoints and LoRA support.
It supports two primary generation modes through separate checkpoints: text-to-video (T2V-A14B) and image-to-video (I2V-A14B). Both support native 480p and 720p generation, with text-guided control over subject appearance, movement, camera behavior and cinematic composition.
Wan 2.2 A14B uses a Mixture-of-Experts (MoE) diffusion architecture with approximately 27 billion total parameters and 14 billion active per denoising step. HexGrid.cloud offers preconfigured 720p video generation up to 5 seconds, with customizable workflows for additional controls and processing.
What people use it for
Cinematic text-to-video
Generate cinematic video scenes from text prompts, controlling subjects, actions, environments, lighting, composition and camera movement through detailed natural-language descriptions.
Image animation
Transform static images into moving videos while preserving their initial composition, subject appearance and artistic style. Guide character actions, environmental motion and camera movement through text prompts.
Product and creative videos
Create short product showcases, promotional visuals, animated illustrations and cinematic social media content from text descriptions or reference images.
Custom styles and characters
Use custom LoRAs to generate videos with specific characters, visual aesthetics, artistic styles or motion characteristics. Build reusable creative workflows for branded content and consistent visual storytelling.
Strengths
- Native 480p and 720p video generation for text-to-video and image-to-video
- Mixture-of-Experts architecture with approximately 27B total parameters and 14B active parameters per denoising step
- Two specialized diffusion experts for high-noise scene structure and low-noise detail refinement
- Cinematic visual quality with strong control over lighting, composition, color grading and camera movement
- Improved complex motion generation, including human actions, facial expressions and dynamic scene interactions
- Image-conditioned generation that uses a reference image to guide the initial appearance and composition
- Support for landscape and portrait video generation, with image-to-video aspect ratios guided by the input image
- Configurable inference steps, guidance scales, frame counts and random seeds through custom workflows
- Optional prompt expansion to enhance scene descriptions and improve generation quality
- Custom LoRA support for specialized characters, artistic styles and visual consistency
- Open-weight architecture supporting fine-tuning, quantization and private deployment
- Apache 2.0 license permitting commercial use, modification and redistribution under its terms
Limitations
- Native generation is limited to 480p and 720p; 1080p and 4K output require external upscaling or additional processing.
- HexGrid.cloud's preconfigured Wan 2.2 workflow supports videos up to 5 seconds. Longer generation requires custom workflows or additional processing.
- Video duration depends on frame count and frame rate rather than a fixed duration selector.
- The base T2V-A14B and I2V-A14B checkpoints generate silent video without native synchronized dialogue, sound effects or music.
- Audio-driven generation requires the separate Wan2.2-S2V-14B model, while advanced character animation and replacement require Wan2.2-Animate-14B.
- Text-to-video and image-to-video use separate A14B checkpoints rather than a unified generation checkpoint.
- Native first-and-last-frame conditioning is not supported by the standard A14B checkpoints and requires additional models or specialized workflows.
- Complex scenes with multiple interacting characters, precise hand movements or extended motion can exhibit temporal inconsistencies and visual artifacts.
- The official A14B single-GPU inference examples recommend at least 80 GB of VRAM; quantization and model offloading can reduce memory requirements.
- Higher inference-step counts increase generation time, while quantization and acceleration techniques may introduce quality trade-offs.
Specifications
Generation, input and serving details.
ProviderQwenFamilyWan-2.2CategoryvideoWorkflows3
Generation
- Modes
- Text-to-video, image-to-video
- Variants
- T2V-A14B, I2V-A14B
- Architecture
- Diffusion Transformer with Mixture-of-Experts (MoE)
- Parameters
- Approximately 27B total, 14B active per denoising step
- Resolution
- 480p, 720p native
- Default resolution
- 720p on HexGrid.cloud
- Duration
- Up to 5 seconds in HexGrid.cloud's preconfigured workflow; configurable through custom workflows
- Frame count
- Configurable; 81 frames is a common approximately 5-second configuration
- Frame rates
- Typically 16 FPS for native A14B generation; higher playback frame rates possible through interpolation
- Default frame rate
- 16 FPS in standard A14B inference configurations
- Aspect ratios
- Landscape and portrait; image-to-video can follow the source image's aspect ratio
- Audio generation
- Not supported natively by T2V-A14B or I2V-A14B
- Multi-shot
- No dedicated native multi-shot timeline control in the base A14B generation workflows
Inputs and Controls
- Text prompt
- Required for text-to-video; optional for image-to-video when supported prompt expansion is used
- Start image
- Required for image-to-video
- End image
- Not natively supported by the standard A14B checkpoints
- Audio input
- Not supported by the base T2V-A14B and I2V-A14B models
- Negative prompt
- Supported for excluding unwanted subjects, artifacts, visual characteristics and motion
- Camera motion
- Controlled through natural-language prompts, including pan, tilt, zoom, dolly, tracking and orbital movements
- Inference steps
- Configurable through the generation workflow
- Guidance scale
- Configurable classifier-free guidance for controlling prompt adherence
- Expert guidance
- High-noise and low-noise diffusion stages can be configured separately in compatible workflows
- Random seed
- Optional integer seed for reproducible generation
- Prompt expansion
- Optional LLM-based enhancement of text prompts
- Frame interpolation
- Optional through compatible ComfyUI interpolation workflows
- LoRA support
- Custom LoRAs supported through compatible ComfyUI workflows
- Quantization
- FP8 and other optimized precision configurations available through compatible runtimes
- Model offloading
- Supported to reduce GPU memory requirements
Serving
- Provider
- Alibaba / Wan Team
- API platform
- HexGrid.cloud
- Deployment
- Private GPU deployment with preconfigured ComfyUI workflows
- Generation workflows
- Text-to-video and image-to-video using separate A14B checkpoints
- Output format
- MP4 video
- Model availability
- Open weights and private hosted deployment
- Model license
- Apache 2.0
- Fine-tuning
- LoRA training and custom model adaptation supported
- GPU requirements
- 80 GB VRAM recommended for official single-GPU A14B inference; optimized deployments may use less
- Infrastructure
- Dedicated GPUs with private model execution and customizable workflows
- Customization
- Custom checkpoints, LoRAs, ComfyUI nodes, inference settings and post-processing workflows
- Performance
- HexGrid.cloud advertises up to 80 clips per hour on H100 SXM for its preconfigured Wan 2.2 workflow; actual throughput varies with configuration