Skip to content

Qwen / video model

Wan 2.2 5B

Explore Wan 2.2 5B, its capabilities, use cases and ComfyUI workflows for video generation on HexGrid Cloud.

Wan-2.22 workflowsExplore workflows ↓

Deploy Wan 2.2 5B 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 5B. Explore the ComfyUI setup, workflow instructions and generation examples.

Text → Video

Explore workflow
02Workflow guide

Image to Video

Animate an input image into video with Wan 2.2 5B. Explore the ComfyUI setup, workflow instructions and generation examples.

Text + Image → Video

Explore workflow

About Wan 2.2 5B

Wan 2.2 5B is Alibaba's efficient, open-weight video generation model, deployable on HexGrid.cloud with dedicated GPU infrastructure, configurable inference workflows, custom checkpoints and LoRA support.

It supports two primary generation modes within a single unified checkpoint: text-to-video and image-to-video. Unlike the larger Wan 2.2 A14B models, which use a Mixture-of-Experts architecture, Wan 2.2 TI2V-5B uses a compact 5-billion-parameter dense Diffusion Transformer optimized for efficient high-definition video generation.

Wan 2.2 5B generates videos at 720p-class resolution and 24 FPS, using an advanced Wan2.2-VAE with 4× temporal and 16×16 spatial compression. Its default configuration produces approximately 5-second clips, while configurable frame counts, sampling parameters and optimized inference enable flexible generation workflows on GPUs with as little as 24 GB of VRAM.

What people use it for

Cinematic text-to-video

Generate high-definition cinematic scenes from detailed text prompts, controlling characters, actions, environments, lighting, camera movements and visual composition through natural-language descriptions.

Image animation

Transform static images into moving videos while preserving the source image's appearance, composition and artistic style. Guide character actions, environmental motion and camera movement through text prompts.

Product and advertising videos

Create product showcases, promotional animations, social media content and branded visuals from text descriptions or reference images, with cinematic lighting, smooth camera movement and dynamic compositions.

Creative animation and storytelling

Animate illustrations, concept art, digital characters and stylized environments into short video sequences. Generate scenes with expressive subject movements, cinematic framing and consistent visual aesthetics.

Rapid video prototyping

Quickly explore cinematic concepts, camera movements, scene compositions and animation ideas using a computationally efficient model suitable for iterative creative workflows.

Strengths

  • Native 720p-class video generation at 24 FPS for both text-to-video and image-to-video
  • Unified 5B-parameter checkpoint supporting both text-conditioned and image-conditioned video generation
  • Compact dense Diffusion Transformer architecture optimized for efficient inference
  • Advanced Wan2.2-VAE with 4× temporal and 16×16 spatial compression for efficient high-definition video generation
  • Additional latent patchification achieving an effective 4×32×32 token compression ratio
  • High-quality cinematic aesthetics with detailed control over lighting, composition, contrast, color grading and camera movement
  • Improved complex motion generation, including character movements, dynamic actions and environmental interactions
  • Image-conditioned generation that preserves source-image appearance while introducing natural motion
  • Support for landscape and portrait video generation, with image-to-video outputs adapting to the input image's aspect ratio
  • Configurable frame counts, diffusion steps, guidance scales, sampling schedules and random seeds
  • Optional LLM-based prompt expansion for enriching scene descriptions and improving generation quality
  • Efficient single-GPU inference with model offloading, supporting GPUs with approximately 24 GB of VRAM
  • Optional multi-GPU inference using PyTorch FSDP and DeepSpeed Ulysses parallelism
  • Custom LoRA support for specialized characters, artistic styles and creative workflows through compatible implementations
  • Open-weight architecture supporting fine-tuning, quantization, custom inference pipelines and private deployment
  • Apache 2.0 license permitting commercial use, modification and redistribution under its terms

Limitations

  • Native generation is designed for 720p-class output, specifically 1280×704 or 704×1280 in standard configurations. Higher resolutions such as 1080p and 4K require external upscaling or additional processing.
  • The default generation length is approximately 5 seconds at 121 frames and 24 FPS. Longer videos require increased frame counts, which raise memory consumption and processing time.
  • Generated frame counts must follow the 4n+1 format, such as 49, 81 or 121 frames, to satisfy the model's temporal compression requirements.
  • Native frame rate is 24 FPS. Higher playback frame rates require external frame interpolation or compatible post-processing workflows.
  • The base TI2V-5B model generates silent videos without native synchronized dialogue, environmental audio, sound effects or music.
  • Audio-driven video generation requires the separate Wan2.2-S2V-14B checkpoint, while character animation and replacement require Wan2.2-Animate-14B.
  • Native first-and-last-frame conditioning is not supported by the standard TI2V-5B checkpoint and requires specialized models or additional workflows.
  • The base model does not provide dedicated multi-shot timeline controls or native long-form scene sequencing.
  • Complex multi-character interactions, precise hand movements, rapid camera changes and extended actions may exhibit temporal inconsistencies or visual artifacts.
  • The compact 5B architecture may produce less fine-grained detail and weaker performance on highly complex scenes compared with larger Wan 2.2 A14B models.
  • Single-GPU inference on 24 GB GPUs requires memory optimizations such as model offloading, reduced precision and CPU-based text encoding.
  • Higher diffusion-step counts improve generation quality at the cost of increased inference time, while aggressive quantization or acceleration may introduce quality trade-offs.

Specifications

Generation, input and serving details.

ProviderQwenFamilyWan-2.2CategoryvideoWorkflows2

Generation

Modes
Text-to-video, image-to-video
Model
Wan2.2-TI2V-5B
Architecture
Dense Diffusion Transformer with unified text-to-video and image-to-video conditioning
Model parameters
5 billion
Text encoder
UMT5-XXL
Resolution
720p-class native generation
Default resolution
1280×704 landscape or 704×1280 portrait
Duration
Approximately 5 seconds by default; configurable through frame count
Default frame count
121 frames
Frame count
Configurable; must follow 4n+1
Frame rates
24 FPS native; higher playback FPS possible through interpolation
Default frame rate
24 FPS
Aspect ratios
Landscape and portrait; image-to-video can follow the source image's aspect ratio
Audio generation
Not supported natively; generates silent video
Multi-shot
No dedicated native multi-shot timeline control
Video decoder
Wan2.2-VAE with high-compression video reconstruction
VAE compression
4× temporal and 16×16 spatial compression
Effective token compression
4×32×32 with additional patchification
Generation framework
Flow-matching diffusion

Inputs and Controls

Text prompt
Required for standard text-to-video; used to guide motion and scene behavior in image-to-video
Start image
Required for image-to-video
End image
Not supported natively by the standard TI2V-5B checkpoint
Audio input
Not supported by the base TI2V-5B model
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; default 50
Guidance scale
Configurable classifier-free guidance; default 5.0
Sampling shift
Configurable flow-matching shift factor; default 5.0
Sampling solver
UniPC or DPM++; UniPC is the default
Random seed
Optional integer seed for reproducible generation; random seed used by default
Prompt expansion
Optional LLM-based enhancement using compatible local or external text models
Prompt expansion languages
English and Chinese
Frame interpolation
Optional through compatible ComfyUI interpolation workflows
Image conditioning
Single-image conditioning with text-guided motion generation
Aspect ratio adaptation
Image-to-video output follows the source image's aspect ratio within the configured resolution area
LoRA support
Custom LoRAs supported through compatible inference workflows
Model precision
BF16 by default; optimized FP8 and other precision configurations available through compatible runtimes
Model offloading
Supported to reduce GPU memory requirements
CPU text encoding
Supported through CPU placement of the UMT5 text encoder
Multi-GPU inference
Supported through PyTorch FSDP and DeepSpeed Ulysses

Serving

Provider
Alibaba / Wan Team
API platform
HexGrid.cloud
Deployment
Private GPU deployment with configurable inference workflows
Generation workflows
Text-to-video and image-to-video using a unified TI2V-5B checkpoint
Model checkpoint
Wan2.2-TI2V-5B
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 through compatible training frameworks
Model architecture
5B dense Diffusion Transformer with high-compression Wan2.2-VAE
Inference frameworks
Native Wan inference, ComfyUI and Diffusers
Infrastructure
Dedicated GPU infrastructure with private model execution and configurable inference pipelines
GPU requirements
Minimum approximately 24 GB VRAM with model offloading and CPU text encoding; higher-memory GPUs recommended for faster inference
Recommended GPU
RTX 4090 or higher with optimized memory settings; 80 GB-class GPUs can run without certain offloading optimizations
Parallel inference
Multi-GPU execution supported through FSDP and DeepSpeed Ulysses
Customization
Custom checkpoints, LoRAs, inference parameters, prompt processing, frame counts and post-processing workflows
Optimization
Model offloading, reduced-precision inference, CPU text encoding and optional quantization
Performance
Optimized for efficient 720p video generation; actual throughput depends on GPU hardware, frame count, diffusion steps and memory configuration

Explore related models