Wan 2.2 5B - Image to Video
ComfyUI in the Cloud
Deploy Wan 2.2 5B - Image to Video, from
$0.60 / hour
Wan 2.2 TI2V 5B ComfyUI Workflow
A ComfyUI video generation workflow configuration powered by Wan 2.2 TI2V 5B, featuring an FP16 diffusion model, UMT5 XXL FP8 text conditioning, and a dedicated Wan 2.2 variational autoencoder for video latent processing.
Workflow Overview
This workflow utilizes the Wan 2.2 TI2V 5B model configuration for AI video generation within ComfyUI. It combines a 5-billion-parameter diffusion model, UMT5 XXL text conditioning, and a dedicated Wan 2.2 VAE for processing video representations.
The TI2V model is designed for text and image-conditioned video generation. Depending on the surrounding workflow, compatible conditioning nodes can provide text instructions and visual information to guide the video generation process.
The primary diffusion model uses FP16 precision, while the UMT5 XXL text encoder uses an FP8 E4M3FN-scaled configuration. The lower-precision text encoder helps reduce memory requirements compared with higher-precision text encoder implementations.
The displayed workflow section contains three model-loading nodes: Load Diffusion Model, Load CLIP, and Load VAE. These nodes provide the primary components required by compatible video generation, conditioning, and decoding nodes. Sampling parameters and output settings are not visible in the screenshot.
Model Configuration
Diffusion Model
Wan 2.2
Model Variant
TI2V 5B
Model Precision
FP16
Text Encoder
UMT5 XXL
Text Encoder Precision
FP8 Scaled
CLIP Type
wan
Video Autoencoder
Wan 2.2 VAE
Model Loading Nodes
3
Included Nodes
Load Diffusion Model
FP16 / 5Bwan2.2_ti2v_5B_fp16
The Load Diffusion Model node loads the Wan 2.2 TI2V 5B model used for video generation. This model is responsible for the denoising process that produces video latent representations. The TI2V designation identifies the model as a text/image-to-video variant. It can be used with compatible conditioning nodes to generate video sequences guided by text instructions and visual inputs.
The selected model uses FP16 precision, representing model weights with 16-bit floating-point values. The node exposes a MODEL output that can be connected to compatible sampling and video generation nodes.
Load CLIP / Text Encoder
UMT5 XXL / FP8umt5_xxl_fp8_e4m3fn_scaled
The Load CLIP node loads the UMT5 XXL text conditioning model used by the Wan video generation architecture. UMT5 XXL processes natural-language instructions and converts them into conditioning representations. These representations provide guidance to the diffusion model during video generation.
The selected text encoder uses an FP8 E4M3FN-scaled configuration. This quantized representation helps reduce memory consumption compared with higher-precision text encoders. The node uses the wan model type and exposes a CLIP output for connection to compatible text conditioning nodes.
Load VAE
Wan 2.2 VAEwan2.2_vae
The Load VAE node loads the dedicated Wan 2.2 variational autoencoder. The VAE provides the encoding and decoding components used for video latent representations. During decoding, the autoencoder converts generated latent sequences into visible video frames.
This component forms an important part of the video generation pipeline. The VAE can be connected to compatible video encoding or decoding nodes within ComfyUI. Its output can support subsequent video preview, processing, or export operations.
Text and Image-to-Video Architecture
The Wan 2.2 TI2V 5B model is designed for video generation using text and image-based conditioning. Depending on the connected workflow nodes, the model can be used for text-guided video generation and image-conditioned video generation.
Text Conditioning
umt5_xxl_fp8_e4m3fn_scaled
Text instructions are processed through the UMT5 XXL encoder. Prompts can describe subjects, actions, camera movements, environments, lighting, and visual styles. The resulting conditioning representations guide the video generation model.
Image Conditioning
Compatible Image Input Nodes
Image-conditioned generation can use an existing image as visual guidance. The reference image may provide information about subjects, composition, appearance, and scene structure. Compatible image conditioning nodes prepare this information for the video generation process.
The screenshot shows the model-loading configuration rather than the complete video generation pipeline. The exact text and image conditioning connections are not visible. Image conditioning behavior depends on the connected nodes and the complete ComfyUI workflow.
Diffusion Model Configuration
The Wan 2.2 TI2V 5B model provides the primary generative component of the workflow. It performs the iterative denoising process used to generate video latent representations.
| Diffusion Model | wan2.2_ti2v_5B_fp16 |
| Model Architecture | Wan 2.2 TI2V |
| Parameter Count | 5B |
| Model Precision | FP16 |
| Node Type | Load Diffusion Model |
| Node Output | MODEL |
The diffusion model is loaded separately from the text encoder and VAE. This modular configuration allows individual model components to be connected to compatible processing nodes within the larger workflow.
Text Encoder Configuration
The workflow uses UMT5 XXL for text conditioning. This model converts natural-language instructions into representations that guide video generation.
| Text Encoder | umt5_xxl_fp8_e4m3fn_scaled |
| Encoder Architecture | UMT5 XXL |
| Precision | FP8 E4M3FN Scaled |
| CLIP Type | wan |
| Node Output | CLIP |
The FP8-scaled configuration helps reduce text encoder memory consumption. The selected CLIP type is set to wan, allowing the encoder to be loaded for compatible Wan conditioning nodes. The encoded text information can then be supplied to the video generation pipeline.
Video VAE Configuration
The workflow includes the dedicated Wan 2.2 VAE. This component provides the autoencoder functionality required for video latent processing.
| VAE Model | wan2.2_vae |
| Architecture | Wan 2.2 |
| Node Type | Load VAE |
| Node Output | VAE |
| Function | Video Latent Encoding / Decoding |
After the diffusion model generates video latent representations, compatible VAE decoding nodes can convert these representations into visible video frames. The decoded frames can then be passed to video combination, preview, or export nodes. The screenshot does not specify the VAE weight precision.
Model Loading and Node Connections
The displayed workflow contains three model-loading nodes. Each node loads a separate component and exposes an output that can be used by other ComfyUI nodes.
Load Diffusion Model → MODEL
Loads the Wan 2.2 TI2V 5B diffusion model. The MODEL output can be connected to compatible video sampling and generation nodes. This component handles the primary denoising process.
Load CLIP → CLIP
Loads the UMT5 XXL text encoder. The CLIP output can be connected to compatible text encoding and prompt conditioning nodes. This component provides text representations for video generation.
Load VAE → VAE
Loads the Wan 2.2 video autoencoder. The VAE output can be connected to compatible latent encoding and decoding nodes. This component supports conversion between video frames and latent representations.
Surrounding Nodes and Workflow Integration
The displayed Wan 2.2 TI2V 5B workflow section prepares the primary model components required for video generation. The diffusion model, text encoder, and VAE are loaded independently and can be connected to compatible downstream workflow components.
In a typical text-guided generation workflow, a text prompt is supplied through a prompt input node. The UMT5 XXL encoder processes the instructions and produces conditioning representations. These representations can be supplied to the generation pipeline.
For image-conditioned video generation, an input image may be supplied through a Load Image node or another upstream image source. Compatible conditioning nodes can prepare the reference image for use by the video generation model.
The Wan 2.2 TI2V 5B diffusion model performs the denoising process using the supplied conditioning information and sampling parameters. The sampling nodes determine the generation schedule, noise initialization, and other inference parameters.
After sampling, the generated latent video representations can be passed to the Wan 2.2 VAE for decoding. Compatible decoding nodes convert the latent representations into visible video frames.
The resulting frames can be connected to video preview, combination, or export nodes. Exact prompt connections, image inputs, resolution, duration, frame rate, sampling settings, and output nodes are not visible in the screenshot.
Workflow Highlights
- ✓ Wan 2.2 TI2V 5B: Uses the Wan 2.2 TI2V 5B diffusion model for AI video generation.
- ✓ FP16 Precision: Uses a 16-bit floating-point diffusion model configuration.
- ✓ 5B Model Architecture: Includes a 5-billion-parameter generation model.
- ✓ UMT5 XXL Conditioning: Provides text conditioning for Wan video generation.
- ✓ FP8 Text Encoder: Uses an FP8 E4M3FN-scaled UMT5 XXL configuration to help reduce memory requirements.
- ✓ Text-Guided Generation: Supports compatible text conditioning workflows for generating video from natural-language instructions.
- ✓ Image Conditioning Support: The TI2V model can be integrated with compatible image conditioning workflows for image-guided video generation.
- ✓ Wan 2.2 VAE: Includes a dedicated video autoencoder for latent encoding and decoding.
- ✓ Modular Model Loading: Uses separate Load Diffusion Model, Load CLIP, and Load VAE nodes.
- ✓ Workflow Integration: Exposes MODEL, CLIP, and VAE outputs for connection to compatible ComfyUI generation nodes.
Workflow Summary
This Wan 2.2 TI2V 5B ComfyUI workflow configuration combines an FP16 diffusion model with UMT5 XXL text conditioning and the dedicated Wan 2.2 VAE. The diffusion model provides the primary video generation component, while the text encoder processes natural-language instructions.
The UMT5 XXL text encoder uses an FP8 E4M3FN-scaled configuration with the CLIP type set to wan. The Wan 2.2 VAE supports video latent processing and decoding within compatible workflow nodes.
The TI2V model can be used in compatible text-guided and image-conditioned video generation pipelines. The displayed section contains three model-loading nodes, exposing MODEL, CLIP, and VAE outputs.
This modular configuration provides the primary model components needed by a larger Wan 2.2 video generation workflow. Additional conditioning, sampling, video dimensions, frame rate, duration, and final output handling depend on the surrounding ComfyUI nodes.