Wan 2.2 14B - First Last Frame to Video
ComfyUI in the Cloud
Deploy Wan 2.2 14B - First Last Frame to Video, from
$0.60 / hour
Wan 2.2 Image-to-Video ComfyUI Workflow
A ComfyUI image-to-video generation workflow powered by Wan 2.2, featuring dual 14B diffusion models, FP8-scaled precision, UMT5 XXL text conditioning, Wan 2.1 VAE decoding, and optional LightX2V Lightning LoRA acceleration.
Workflow Overview
This workflow utilizes the Wan 2.2 image-to-video architecture for AI-powered video generation within ComfyUI. It combines separate high-noise and low-noise diffusion models, UMT5 XXL text conditioning, and the Wan 2.1 VAE for video latent decoding.
Unlike a single-model diffusion workflow, Wan 2.2 uses a dual-stage denoising architecture. The high-noise model processes the earlier stages of generation, where the latent representation contains higher noise levels. The low-noise model handles later denoising stages, refining the generated representation as the noise level decreases.
Both diffusion models use FP8-scaled configurations to help reduce memory requirements compared with higher-precision model weights. The workflow also includes separate Lightning LoRA models designed for the high-noise and low-noise stages.
The displayed configuration uses a video resolution of 640 × 640 pixels, a duration value of 5.0, and a noise seed of 264244520398999. Turbo Mode is disabled, meaning the optional Lightning acceleration configuration is not enabled in the displayed workflow.
Generation Settings
Width
640
Height
640
Video Duration
5.0
Noise Seed
264244520398999
Model Architecture
Dual 14B
Model Precision
FP8 Scaled
Turbo Mode
Disabled
Resolution
640 × 640
Model Configuration
High-Noise Diffusion Model
14B / FP8wan2.2_i2v_high_noise_14B_fp8_scaled
The high-noise diffusion model handles the initial stages of the video generation process. During these stages, the latent representation contains relatively high levels of noise. The model performs early denoising operations, helping establish the overall visual structure, composition, and motion representation. The selected model contains 14 billion parameters and uses an FP8-scaled configuration. FP8 precision helps reduce model memory requirements compared with higher-precision weights. This model forms the first major component of the dual-stage Wan 2.2 generation architecture.
Low-Noise Diffusion Model
14B / FP8wan2.2_i2v_low_noise_14B_fp8_scaled
The low-noise diffusion model handles the later stages of the generation process. After the high-noise model establishes the initial latent representation, the low-noise model continues denoising as the noise level decreases. This stage focuses on refining the generated video representation. It contributes to the development of finer visual details and the final latent structure. The model uses a 14B architecture with FP8-scaled precision. Together, the high-noise and low-noise models form the dual-stage generation configuration used by Wan 2.2.
Text Conditioning Model
UMT5 XXLumt5_xxl_fp8_e4m3fn_scaled
The UMT5 XXL model provides text conditioning for the Wan 2.2 image-to-video generation process. It processes natural-language instructions and converts them into conditioning representations used by the diffusion models. These representations help guide the generation process according to the supplied instructions. The selected text encoder uses an FP8 E4M3FN-scaled configuration. This lower-precision representation helps reduce memory consumption compared with higher-precision text encoder implementations.
Video Variational Autoencoder (VAE)
Wan 2.1 VAEwan_2.1_vae
The workflow uses the Wan 2.1 VAE for handling video latent representations. During decoding, the VAE converts generated latent representations into visible video frames. This component provides the decoding stage required to transform the generated latent sequence into pixel-space output. The decoded frames can then be passed to compatible video preview, processing, or export nodes within ComfyUI.
Dual-Stage Diffusion Architecture
Wan 2.2 uses separate high-noise and low-noise diffusion models rather than relying on a single model for the entire denoising process. This allows different model components to handle different noise ranges during generation.
Stage 1: High-Noise Processing
wan2.2_i2v_high_noise_14B_fp8_scaled
Handles the initial denoising stages when the latent representation contains higher noise levels. This stage helps establish the initial visual and motion structure of the generated video.
Stage 2: Low-Noise Processing
wan2.2_i2v_low_noise_14B_fp8_scaled
Handles the later denoising stages when the latent representation contains lower noise levels. This stage refines the generated representation and contributes to the final visual details.
The transition between high-noise and low-noise processing is controlled by the internal sampling configuration. The exact switching point and sampling schedule are not visible in the screenshot. Both models are selected as part of the displayed Wan 2.2 configuration.
LightX2V Lightning LoRA Configuration
The workflow includes separate LightX2V Lightning LoRA models for the high-noise and low-noise diffusion stages. These LoRAs are designed for accelerated image-to-video generation using a 4-step configuration.
Separate Lightning LoRAs are provided because Wan 2.2 uses distinct diffusion models for different noise ranges. Each LoRA is associated with its corresponding diffusion stage.
High-Noise Lightning LoRA
4-Step / v1wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise
This Lightning LoRA is designed for the high-noise diffusion model. It provides an acceleration configuration intended for the initial denoising stage. The selected model is identified as a 4-step Lightning variant, version 1. It is associated with the high-noise model when the workflow's Turbo Mode is enabled.
Low-Noise Lightning LoRA
4-Step / v1wan2.2_i2v_lightx2v_4steps_lora_v1_low_noise
This Lightning LoRA is designed for the low-noise diffusion model. It provides an acceleration configuration intended for the later denoising stage. The selected model is identified as a 4-step Lightning variant, version 1. It is associated with the low-noise model when the workflow's Turbo Mode is enabled.
Turbo Mode Configuration
The workflow includes an optional Turbo Mode for accelerated video generation. The selected Lightning LoRA models are designed for reduced-step inference using their respective high-noise and low-noise diffusion stages.
| Turbo Mode | Disabled |
| Lightning Variant | LightX2V |
| Lightning Step Designation | 4-Step |
| Lightning Version | v1 |
| High-Noise LoRA | Selected |
| Low-Noise LoRA | Selected |
Turbo Mode is disabled in the displayed workflow configuration. The Lightning LoRA models remain selected, but the optional acceleration mode is not enabled. The 4-step designation comes from the selected LoRA model names. The actual sampling step count is not visible in the screenshot and should not be assumed from the LoRA names alone.
Video Generation Configuration
The workflow includes dedicated controls for video width, height, duration, and noise seed configuration. These parameters determine the requested spatial and temporal characteristics of generated video.
Video Width
640 px
Controls the horizontal dimension of the generated video. The displayed width is set to 640 pixels.
Video Height
640 px
Controls the vertical dimension of the generated video. The displayed height is set to 640 pixels.
Video Duration
5.0
Defines the configured duration of the generated video. In a seconds-based duration configuration, this represents five seconds.
The selected dimensions produce a square 640 × 640 configuration with a 1:1 aspect ratio. The duration value is set to 5.0. The actual frame count and frame rate depend on the internal generation nodes and are not visible in the screenshot.
Noise Seed Configuration
The workflow includes a configurable noise seed that controls random initialization during video generation. The displayed seed value is 264244520398999.
| Noise Seed | 264244520398999 |
| Seed Control | Configurable |
| Randomization | Available |
A fixed seed can help reproduce results when the reference image, prompt, model configuration, and other generation parameters remain unchanged. Changing the seed modifies random initialization and may produce different video variations.
Additional Parameters
| Workflow Type | Image-to-Video |
| Width | 640 |
| Height | 640 |
| Resolution | 640 × 640 |
| Aspect Ratio | 1:1 |
| Duration | 5.0 |
| Noise Seed | 264244520398999 |
| High-Noise Model | Wan 2.2 I2V 14B FP8 |
| Low-Noise Model | Wan 2.2 I2V 14B FP8 |
| Text Encoder | UMT5 XXL FP8 |
| VAE | Wan 2.1 VAE |
| Turbo Mode | Disabled |
| Lightning LoRA | LightX2V 4-Step v1 |
| Sampling Steps | Not visible in the screenshot |
Surrounding Nodes and Workflow Integration
The Wan 2.2 image-to-video subgraph consolidates high-noise and low-noise diffusion models, text conditioning, video decoding, generation dimensions, noise initialization, and optional Lightning acceleration settings.
In a typical image-to-video workflow, an input image provides the initial visual reference for video generation. The image may originate from a Load Image node, another image generation workflow, or an upstream image processing component. Compatible image conditioning nodes prepare the reference information for video generation.
Text instructions can be processed through the UMT5 XXL conditioning model. These instructions may describe camera movement, subject motion, environmental changes, and other desired video characteristics. The resulting conditioning information helps guide the diffusion process.
The high-noise diffusion model handles the initial generation stage. It processes the latent representation during the higher-noise portion of the denoising schedule. The low-noise diffusion model then handles the later denoising stage, refining the video representation.
The workflow includes separate Lightning LoRA configurations for both diffusion stages. When Turbo Mode is enabled, compatible workflow logic can apply the high-noise Lightning LoRA to the high-noise model and the low-noise Lightning LoRA to the low-noise model. Turbo Mode is disabled in the displayed configuration.
The width and height parameters determine the configured spatial dimensions of the generated video. The duration parameter defines the requested sequence length, while the noise seed controls random initialization.
After the diffusion process, the selected Wan 2.1 VAE decodes the generated latent representations into visible video frames. These frames can then be passed to compatible video combination, preview, or export nodes.
The screenshot shows the subgraph's model selections and generation parameters. Exact image input connections, prompt nodes, sampler settings, frame rate, internal denoising schedule, and final output nodes are not visible in the supplied screenshot.
Workflow Highlights
- ✓ Wan 2.2 Image-to-Video: Uses the Wan 2.2 architecture for AI video generation from image-based conditioning.
- ✓ Dual 14B Models: Includes separate high-noise and low-noise diffusion models.
- ✓ FP8 Quantization: Uses FP8-scaled diffusion model configurations to help reduce memory requirements.
- ✓ High-Noise Processing: Uses a dedicated diffusion model for the initial denoising stages.
- ✓ Low-Noise Refinement: Uses a separate diffusion model for the later denoising stages.
- ✓ UMT5 XXL Conditioning: Provides text conditioning using the FP8 E4M3FN-scaled encoder.
- ✓ Wan 2.1 VAE: Includes a dedicated VAE for decoding video latent representations into frames.
- ✓ Lightning LoRA Support: Includes separate LightX2V 4-step LoRA models for high-noise and low-noise stages.
- ✓ Optional Turbo Mode: Provides an acceleration option using the selected Lightning LoRAs, currently disabled.
- ✓ 640 × 640 Resolution: Configured for square video generation with a 1:1 aspect ratio.
- ✓ Configurable Duration: Includes a duration parameter currently set to 5.0.
- ✓ Seed-Based Generation: Supports configurable noise initialization with the displayed seed set to 264244520398999.
- ✓ Reusable Subgraph: Organizes dual-stage diffusion, model selection, Lightning LoRAs, and video parameters into a reusable ComfyUI component.
Workflow Summary
This Wan 2.2 image-to-video ComfyUI workflow combines two 14B FP8-scaled diffusion models for high-noise and low-noise denoising stages. It uses UMT5 XXL text conditioning and the Wan 2.1 VAE for processing and decoding video latent representations.
The workflow is configured with a resolution of 640 × 640 pixels, a duration value of 5.0, and a noise seed of 264244520398999. The dual-stage diffusion configuration separates initial denoising from later refinement operations.
The configuration also includes separate LightX2V Lightning LoRA models for the high-noise and low-noise stages. Both selected LoRAs are identified as 4-step version 1 variants. Turbo Mode is currently disabled, so the optional Lightning acceleration configuration is not enabled.
The subgraph organizes model selection, dual-stage diffusion, text conditioning, VAE decoding, generation dimensions, noise initialization, and optional acceleration settings. This configuration can be integrated into larger ComfyUI image-to-video workflows using compatible input, conditioning, sampling, and output nodes.