Video-subtitle-remove (VSR) is a software based on AI technology that removes hard subtitles from videos. It mainly realizes the following functions:
- Lossless resolution removes hard subtitles from the video and generates a caption-removed file
- Fill the area from which subtitle text is removed through a super AI algorithm model (non-adjacent pixel filling and mosaic removal)
- Support custom subtitle positions, only remove subtitles in defined positions (incoming positions)
- Support automatic removal of all text from full video (not passed into location)
- Support multi-selection images to remove watermarked text in batches
Instructions for use:
If you have any usage problems, please discuss in groups,
Directly download the compressed package, decompress it and run it. If it cannot run it, follow the following tutorial and try to install the source code and run it in the conda environment.
Windows GPU version v1.1.0 (GPU):
Baidu Online Disk: windows_gpu_v1.1.0.zip Extraction Code: vsr1
Google Drive: vsr_windows_gpu_v1.1.0.zip
Only for users with Nvidia graphics cards (AMD graphics cards are not available)
common problems
What if the extraction speed is slow
Modifying the parameters in backend/config.py can greatly increase the removal speed
MODE = InpaintMode.STTN #Set to STTN algorithm
STTN_SKIP_DETECTION = True #Skip subtitle detection. After skipping may cause the subtitles to be removed to be missed or the video frames that do not need to be removed to be accidentally injured
What should I do if the video removal effect is not good
Modify the parameters in backend/config.py and try different removal algorithms. Algorithm introduction
InpaintMode.STTN algorithm: It has good effects for live video, is fast, and can skip subtitle detection
InpaintMode.LAMA algorithm: It has the best effect on pictures and good effect on animation videos. The speed is average. Subtitle detection cannot be skipped
InpaintMode. PROPAINTER algorithm: It consumes a lot of video memory, is slow, and has a good effect on videos with very intense exercise
Using STTN algorithm
MODE = InpaintMode.STTN #Set to STTN algorithm
Adjacent frames, adjusting will increase the memory consumption, and the effect will be better
STTN_NEIGHBOR_STRIDE = 10
Referring to the frame length, the adjustment will increase the memory consumption, and the effect will be better
STTN_REFERENCE_LENGTH = 10
Set the maximum number of frames that the STTN algorithm processes simultaneously. The higher the setting, the slower the speed, but the better the effect
Ensure that STTN_MAX_LOAD_NUM is greater than STTN_NEIGHBOR_STRIDE and STTN_REFERENCE_LENGTH
STTN_MAX_LOAD_NUM = 30
Using LAMA algorithm
MODE = InpaintMode.LAMA #Set to STTN algorithm
LAMA_SUPER_FAST = False #Guaranteed effect
If you are not satisfied with the captioning effect of the model, you can check the training method in the design folder, use the code in backend/tools/train to train, and then replace the trained model with the old model
CondaHTTPError
Place the.condarc in the project in the user directory (C:/Users/your user name), and overwrite it if the file already exists in the user directory
7z file decompression error
Solution: Upgrade the 7-zip decompressor to the latest version
4090 Can’t run with cuda 11.7
Solution: Switch to cuda 11.8
If you want to learn more, you can click on the link below the video.
Thank you for watching this video. If you like it, please subscribe and like it. thank
Download address:https://download.pytorch.org/whl/cu118
GitHub：https://github.com/YaoFANGUK/video-subtitle-remover
Video: