How can I remove text from video without blur?

DeepMosaics is specifically designed to remove text from video AI-based by using "semantic segmentation" and "image-to-image translation" . Its core workflow involves:

1. Training/Model Requirement: The tool uses pre-trained AI models to recognize and "clean" specific types of artifacts .
2. Model Limitations: The main pre-trained models are designed for face mosaics or a specific other type of mosaic. There is no pre-trained model specifically for "video text removal" .
3. Custom Training Required: To actually remove text from video AI-powered, you would likely need to train your own AI model using the tool's custom dataset creation scripts, requiring a dataset of text-mosaic videos and a capable NVIDIA GPU (6GB+ memory) .

You can run DeepMosaics either via a pre-built Windows GUI version or from the source code.

Option 1: Pre-built GUI (Windows only)

• Pros: User-friendly interface; no coding required .
• Cons: Lags behind the source code in updates; still requires you to download the correct pre-trained model.

Option 2: Run from Source (All OS)

This is the more flexible method but requires technical setup.

Prerequisites

• Python 3.6+
• FF mpeg 3.4.6
• PyTorch 1.0+
• NVIDIA GPU with CUDA (recommended for video; CPU is very slow)

To Remove a Mosaic (e.g., on a face):

bash

python deepmosaic.py --media_path ./input_video.mp4 --model_path ./pretrained_models/mosaic/clean_face_HD.pth --gpu_id 0

The output will be saved in the ./result folder .

Key Parameter:

• --model_path: You must specify the correct pre-trained model for the artifact you want to remove. The models provided are for faces and a specific type of mosaic, not generic text.

For your specific goal of removing text from video, Video-subtitle-remover (VSR) is a more suitable choice. Unlike DeepMosaics, VSR is specifically designed for this task and comes with pre-trained models for detecting and inpainting subtitles and text-like watermarks.