Wan FLF2V (First-Last Frame Video Generation) is an open-source video generation model developed by the Alibaba Tongyi Wanxiang team. Its open-source license is Apache 2.0. Users only need to provide two images as the starting and ending frames, and the model automatically generates intermediate transition frames, outputting a logically coherent and naturally flowing 720p high-definition video.

Core Technical Highlights

  1. Precise First-Last Frame Control: The matching rate of first and last frames reaches 98%, defining video boundaries through starting and ending scenes, intelligently filling intermediate dynamic changes to achieve scene transitions and object morphing effects.
  2. Stable and Smooth Video Generation: Using CLIP semantic features and cross-attention mechanisms, the video jitter rate is reduced by 37% compared to similar models, ensuring natural and smooth transitions.
  3. Multi-functional Creative Capabilities: Supports dynamic embedding of Chinese and English subtitles, generation of anime/realistic/fantasy and other styles, adapting to different creative needs.
  4. 720p HD Output: Directly generates 1280×720 resolution videos without post-processing, suitable for social media and commercial applications.
  5. Open-source Ecosystem Support: Model weights, code, and training framework are fully open-sourced, supporting deployment on mainstream AI platforms.

Technical Principles and Architecture

  1. DiT Architecture: Based on diffusion models and Diffusion Transformer architecture, combined with Full Attention mechanism to optimize spatiotemporal dependency modeling, ensuring video coherence.
  2. 3D Causal Variational Encoder: Wan-VAE technology compresses HD frames to 1/128 size while retaining subtle dynamic details, significantly reducing memory requirements.
  3. Three-stage Training Strategy: Starting from 480P resolution pre-training, gradually upgrading to 720P, balancing generation quality and computational efficiency through phased optimization.

Related Links

​
Wan2.1 FLF2V 720P ComfyUI Native Workflow Example

We provide only the fp16 version of the model, which may cause difficulties in normal operation under lower VRAM.

Please download the WebP file below, and drag it into ComfyUI to load the corresponding workflow. The workflow has embedded the corresponding model download file information.

Please download the two images below, which we will use as the starting and ending frames of the video

​
2. Manual Model Installation

If corresponding

All models involved in this guide can be found here.

diffusion_models Choose one version based on your hardware conditions

If you have previously tried Wan Video related workflows, you may already have the following files.

Choose one version from Text encoders for download,

VAE

CLIP Vision

File Storage Location

ComfyUI/
├── models/
│   ├── diffusion_models/
│   ├── wan2.1_flf2v_720p_14B_fp16.safetensors
│   ├── text_encoders/
│   │   └─── umt5_xxl_fp8_e4m3fn_scaled.safetensors           # or your chosen version
│   └── vae/
│   │   └──  wan_2.1_vae.safetensors
│   └── clip_vision/
│       └──  clip_vision_h.safetensors

​
3. Complete Workflow Execution Step by Step

  1. Ensure the Load Diffusion Model node has loaded wan2.1_flf2v_720p_14B_fp16.safetensors
  2. Ensure the Load CLIP node has loaded umt5_xxl_fp8_e4m3fn_scaled.safetensors
  3. Ensure the Load VAE node has loaded wan_2.1_vae.safetensors
  4. Ensure the Load CLIP Vision node has loaded clip_vision_h.safetensors
  5. Upload the starting frame to the Start_image node
  6. Upload the ending frame to the End_image node
  7. (Optional) Modify the positive and negative prompts, both Chinese and English are supported
  8. (Optional) Modify the video size in WanFirstLastFrameToVideo, it is recommended to use a small size for initial testing, and then set a larger size after completion to reduce waiting time
  9. Click the Run button, or use the shortcut Ctrl(cmd) + Enter to execute video generation

Was this page helpful?

Suggest edits