zsxkib/keep 🔢✓🖼️❓ → 🖼️

▶️ 108 runs 📅 May 2025 ⚙️ Cog 0.14.9 🔗 GitHub 📄 Paper ⚖️ License
video-restoration video-upscaling

About

🧼Upscales faces in videos look to be clearer and better using KEEP, Kalman-Inspired Feature Propagation for Video Face Super-Resolution🫟

Example Output

Output

Performance Metrics

20.21s Prediction Time
59.19s Total Time
All Input Parameters 
{
  "draw_box": true,
  "input_video": "https://api.replicate.com/v1/files/YmFjMzJiY2YtY2ViMC00MjJjLTg0MTItYWFmYmU4NGYyOGU1",
  "bg_enhancement": false
}
Input Parameters
bg_tile Type: integerDefault: 400Range: 64 - ∞
Tile size for background upsampler (RealESRGAN)
upscale Type: integerDefault: 1Range: 1 - ∞
Upscaling factor for enhancement (RealESRGAN)
draw_box Type: booleanDefault: false
Draw box around detected faces
max_length Type: integerDefault: 20Range: 1 - ∞
Maximum number of frames to process in one batch for KEEP model
has_aligned Type: booleanDefault: false
Set to true if faces in the video are already aligned and cropped (512x512)
input_video (required) Type: string
Input video file
face_upsample Type: booleanDefault: false
Upsample restored faces using RealESRGAN
bg_enhancement Type: booleanDefault: false
Enable background enhancement using RealESRGAN
save_video_fps Type: number
FPS for the output video. If None, uses original FPS.
detection_model Default: retinaface_resnet50
Face detection model to use
only_center_face Type: booleanDefault: true
Process only the center face if multiple faces are detected
Output Schema

Output

Type: string • Format: uri

Example Execution Logs 
Background upsampling: False, Face upsampling: False
Detecting keypoints and smoothing alignment ...
Restoring faces ...
Pasting faces back ...
Saving video ...
All results are saved in /tmp/cog_outputs/current_prediction/tmp65mci3dvdownload_restored.mp4.
Version Details
Version ID
51d5aa9f475690a7cec1f9684efec31f70beca623c27b96dfb604f8223aa437b
Version Created
May 20, 2025
Run on Replicate →