README.md

metadata

title: Video Watermark Remover
emoji: 🦎
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: 4.44.0
python_version: '3.10'
app_file: app.py
pinned: false
license: apache-2.0

Video Watermark Remover

Self-hosted Hugging Face Space for removing static, opaque watermarks from video footage.

Modes

Mode	Model	Speed	Best for
Fast	LaMa (per-frame)	Seconds	Sky, water, foliage — low-frequency backgrounds
Quality	Wan2.1-VACE-14B + lightx2v 4-step distill, FP8 on H200	~2-3 min	Structured or textured backgrounds

How it works

1. Upload a video clip — up to 60 s; the first 15 s is processed at 1080p
2. Paint over the watermark on the first frame using the brush, or:
   • Snap to Rectangle — scribble roughly, click to convert to a clean filled rectangle
   • Clear Mask — start the drawing over
3. Choose Fast or Quality mode
4. Hit Remove Watermark — output is composited back at the source resolution

Crop-inpaint-composite

The pipeline never runs the model on the full 1920×1080 frame:

• The mask determines a tight crop, expanded to a model-friendly resolution with surrounding context
• Only that crop (~7× fewer pixels) is processed
• The result is feather-blended back into the original frame
• All other pixels are byte-identical to the source

V-Log / HDR colour metadata is preserved via FFmpeg flag passthrough (10-bit H.265 output for HDR sources).

Platform

• ZeroGPU (Nvidia H200 MIG slice)
• PyTorch + diffusers ≥0.34
• ffmpeg via packages.txt

Upstream protection

All model files come from a private mirror at JackIsNotInTheBox/Video_Watermark_Remover_Checkpoints:

• LaMa: lama/big-lama.pt — LAMA_MODEL_URL defaults to the mirror, prefetched into torch.hub cache before simple_lama_inpainting can reach for its hardcoded GitHub release URL
• VACE-14B: vace-14b/ — full diffusers package, loaded with local_files_only=True so any cache miss errors loudly instead of silently fetching from upstream HF Hub
• Distill LoRA: loras/wan2.1_t2v_14b_lora_rank64_lightx2v_4step.safetensors — same local_files_only=True enforcement

The Space stays functional even if upstream Wan-AI / lightx2v / GitHub release sources are deleted: at runtime nothing reaches for them.

On the first deploy, ~75 GB of VACE weights are downloaded from the mirror to the persistent cache in a background thread. Fast mode works immediately; Quality mode blocks until prewarm finishes (the UI shows a progress message during the wait).

License

• Pipeline code: Apache 2.0
• LaMa: Apache 2.0
• Wan2.1-VACE-14B: Apache 2.0 (Wan-AI repo)
• lightx2v 4-step distill LoRA: Apache 2.0