What is Wan-Animate?

Wan-Animate is a Wan 2.1-14B-I2V based model specialized for motion transfer to humans and characters.

  • Animation Mode: Moves the input image according to the movement of the reference video.
  • Replacement Mode: Replaces the person in the input video with the person in the reference image.

There are two generation modes, but it is easier to think of Replacement Mode as Animation Mode with added "processing to blend into the background".

Since it is based on Wan 2.1, it can only generate up to 77 frames, but like the Wan 2.1 VACE extension, it has a feature that allows you to generate virtually infinitely long videos by repeatedly extracting the last few frames and generating the continuation.

Required Custom Nodes

Face detection and pose estimation are performed as pre-processing. The following custom nodes are very convenient.

Model Download

Collect Wan-Animate core and common models for the Wan 2.1 series.

Placement example:

📂ComfyUI/
└── 📂models/
    ├── 📂clip_vision/
    │   └── clip_vision_h.safetensors
    ├── 📂diffusion_models/
    │   └── Wan2_2-Animate-14B_fp8_e4m3fn_scaled_KJ.safetensors
    ├── 📂loras/
    │   └── WanAnimate_relight_lora_fp16.safetensors
    ├── 📂text_encoders/
    │   └── umt5_xxl_fp8_e4m3fn_scaled.safetensors
    ├── 📂unet/
    │   └── Wan2.2-Animate-14B-XXXX.gguf      ← Only when using gguf
    └── 📂vae/
        └── wan_2.1_vae.safetensors

Animation Mode

This mode moves the input still image according to the movement of the person in the reference video.

It's quite huge so it might be intimidating, but the base is exactly the same form as Wan 2.1 image2video. Let's go ahead without fear!

wan2_2_animate_animation.json
Deep diveAnimation Mode

Load Wan-Animate Model

  • Load Wan2_2-Animate with Load Diffusion Model.

Decide Generation Resolution

  • Adjust the total number of pixels with Scale Image to Total Pixels according to the input image.
  • Change the value according to your PC specs.
  • Finally, crop the resolution to a multiple of 16.

Input additional information to WanAnimateToVideo node

  • reference_image The still image you want to move.
  • face_video Video with the face part cropped from the reference video. Pose and Face Detection automatically performs face detection by YOLO -> cropping.
  • pose_video Video generating stick figures (key points) from the reference video using ViTPose. Since the skeleton and position are different between the driving video and the image you want to move, adjustments are made by the retarget process.
Outputs
reference_image
reference_image
pose_video(before processing)
output

Replacement Mode

This mode replaces the person in the input video with the person in the reference image.

It adds a mask for inpainting the person and relighting processing to blend into the background to the Animation Mode.

wan2_2_animate_replacement.json
Deep diveReplacement Mode

Add Relight LoRA

  • Add relight LoRA to blend the replaced person into the background.

Padding of Reference Image

  • Since the video is the standard this time, pad the reference image according to the resolution of the video.

Person Mask Generation

  • Pass the person coordinates acquired by Pose and Face Detection to SAM2.1 to generate a mask.
  • Inflate the mask slightly and convert it to a blocky mask like pixel art with the Blockify node to make it a character_mask. If you don't do this, for some reason a thin edge remains on the outline of the generated video.
  • Use a video where the masked part is filled with black with ImageCompositeMasked as background_video.
Outputs
background-pose_video
reference_image
reference_image
output

6-Step Inference (Lightx2v LoRA)

You can reduce sampling steps to 4-6 steps using Distilled LoRA.

I was concerned about degradation when using it with text2video, but with Wan-Animate, since we are not creating a video from scratch, it doesn't bother me much. I want to actively use it.

Model Download

📂ComfyUI/
└── 📂models/
    └── 📂loras/
        └── Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors

Animation Mode (Fast Version)

wan2_2_animate_animation_lightx2v.json
Deep diveAnimation Mode (Fast Version)

Apply LoRA

  • 🟪 Load Lightx2v LoRA with LoraLoaderModelOnly.

  • KSampler Settings

    • steps ... 4-6
    • cfg ... 1.0
Outputs
20steps
6steps

Replacement Mode (Fast Version)

wan2_2_animate_replacement_lightx2v.json

Repeating Process for Long Videos

The base of Wan-Animate is the same as Wan 2.1 I2V, and the upper limit is 77 frames generated in one inference. If you want to create a long video exceeding this, construct it to "repeat the generation many times while inheriting the last few frames".

Since ComfyUI cannot perform loop processing, it will be in the form of connecting almost the same workflow in series one after another.

Allows specific to be frankly not a smart process, and is a part where it yields a step to Kijai's implementation in ComfyUI-WanVideoWrapper.

Animation Mode (Repeat)

wan2_2_animate_animation_lightx2v_repeat.json

At first glance, it looks like a huge workflow, but the only differences from the previous ones are the following two points.

  • video_frame_offset

    • If 77 frames were generated in the first round, face_video and pose_video need to be used from the 78th frame onwards in the second round.
    • If you put the offset frame count in video_frame_offset, it will automatically shift the reference start position of face_video / pose_video.

  • continue_motion_max_frames

    • Set the number of frames to serve as overlap.
    • For example, if length is 77 and continue_motion_max_frames is 5, it uses the last 5 frames from the previous round and generates the remaining new 72 frames.

If you connect this group repeatedly, you can theoretically make a video as long as you want. However, like a copier, the error accumulates little by little.

Replacement Mode (Repeat)

wan2_2_animate_replacement_lightx2v_repeat.json