Skip to content

README.md

Latest commit

 

History

History
117 lines (78 loc) · 4.59 KB

README.md

File metadata and controls

117 lines (78 loc) · 4.59 KB

ComfyUI WAN I2V Control

A user-friendly way to selectively transform parts of your starting image in WAN Image-to-Video generation.

Watch the demo

Watch the demo video

What It Does

This node pack intercepts the conditioning in WAN I2V and uses masking to control which parts of your initial frame get transformed. Instead of the whole starting image being subject to I2V transformation, you can target specific regions - like changing just a character's face while keeping the background intact in that first frame.

Buy Me A Coffee

Use Cases

  • Character transformation - Change a person's appearance in the starting image while preserving the scene
  • Selective regeneration - Fix just the face or hair in your initial frame
  • Multi-person scenes - Target only one person when there are multiple in frame
  • Scene continuity - Take the last frame of a previous I2V clip, regenerate the character's face, then continue to the next video segment

An LTX 2 version is planned if there is sufficient interest.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/shootthesound/comfyui-wan-i2v-control.git
pip install mediapipe

Restart ComfyUI. Example workflows are included in the example_workflows folder.

Quick Start

Try the example workflow first! Load example_workflows/Wan Demo.json -- this is the workflow from the demo video and will give you the best start.

Or build your own:

  1. Add WAN I2V Conditioning Mask Pro node to your workflow
  2. Connect it between your conditioning and the sampler
  3. Enable generate_person_mask
  4. Select what to change: mask_face, mask_hair, mask_body, mask_clothes
  5. Run - only the selected regions will be transformed

Targeting One Person in a Multi-Person Scene

When you have multiple people in frame and only want to change one:

  1. Set ignore_area to "left" or "right"
  2. Set ignore_percent to 0.5 (or adjust as needed)
  3. Only the person on the non-ignored side will be detected and transformed

Mask Options

Built-in Person Detection (MediaPipe):

  • mask_face - Face skin area
  • mask_hair - Hair
  • mask_body - Body/skin
  • mask_clothes - Clothing
  • mask_background - Background only

Face Landmarks (works best on close-ups or 720p+):

  • mask_face_oval, mask_eyes, mask_eyebrows, mask_lips, mask_pupils
  • mask_nose, mask_cheeks, mask_forehead, mask_jaw_chin, mask_ears

Preset Modes (when not using person detection):

  • full, face_focus, top_half, bottom_half, left_half, right_half
  • center, edges, gradient_top, gradient_bottom, soft_vignette

External Inputs:

  • Connect a custom mask input (overrides built-in detection)
  • Connect a depth_map for distance-based masking with depth_threshold

Controls

Option Description
mask_strength Blend between masked and unmasked (0.0-1.0)
grow_mask Expand (positive) or shrink (negative) the mask
feather Soft edge blur
invert_mask Flip which regions get changed
text_strength Adjust prompt influence relative to the image
tint_fill / tint_color Tint masked regions (experimental, results vary)

Drop First Frames Node

Also included is Drop First Frames - a simple but useful utility node for any I2V workflow, not just this pack.

With any I2V generation (including this project), the first couple of frames can sometimes be garbled or weird. This node lets you drop them automatically:

  • Connect your video output to this node
  • Set frames_to_drop (default: 4)
  • Clean output with the bad frames removed

Tips

  • Ghost edges? Try setting feather to around 0.015 - helps blend the edges of dynamic masks.
  • The first couple of frames in I2V can be weird - use the Drop First Frames node
  • Higher resolution (720p+) gives better face landmark detection
  • For best results with face features, use close-up shots
  • You can use LoadImage's MASK output for alpha-based masking from transparent PNGs
  • Experiment with a grow mask feature in postitive or negative values

Credits

Person and face detection powered by MediaPipe.

Author: Peter Neill (ShootTheSound)

Contributing

PRs welcome! Particularly interested in adding support for first frame / last frame / middle frame targeting.

License

MIT