What happened?
Thank you for your hard work in creating such a modern release.
I tested VideoSubtitleRemoverPro v3.18.1 and noticed that all of the available models produce results similar to those shown in the attached image. It seems that these results do not reflect the actual capabilities of the models.
When I use STTN or LaMa in other software, the removed subtitle/text areas are clearly filled in with plausible content (inpainting), resulting in much cleaner output. I'm not sure why your software doesn't seem to be producing the same behavior. I would appreciate it if you could either fix this issue or provide guidance on how to configure the software to achieve the expected results.
I also have another question. Does the LaMa model support an intelligent subtitle region selection feature similar to STTN? From my testing, the automatic subtitle detection (including STTN's automatic detection) is not nearly as accurate as when I manually select the region myself.
In my case, the videos do not contain subtitles—they only contain static text overlays. I have found that LaMa actually removes the text very well, but because it fails to detect the entire text region, some parts remain untouched. Therefore, I would really appreciate it if you could add an option that allows users to manually select the region to be processed when using LaMa, just like with STTN. I believe this would greatly improve the results for videos with fixed text overlays.
Steps to reproduce
1
Version
VideoSubtitleRemoverPro v3.18.1
Surface
GUI
Operating system
Windows 11
GPU
NVIDIA (CUDA)
Support bundle
No response
Media summary
No response
Additional context
No response
What happened?
Thank you for your hard work in creating such a modern release.
I tested VideoSubtitleRemoverPro v3.18.1 and noticed that all of the available models produce results similar to those shown in the attached image. It seems that these results do not reflect the actual capabilities of the models.
When I use STTN or LaMa in other software, the removed subtitle/text areas are clearly filled in with plausible content (inpainting), resulting in much cleaner output. I'm not sure why your software doesn't seem to be producing the same behavior. I would appreciate it if you could either fix this issue or provide guidance on how to configure the software to achieve the expected results.
I also have another question. Does the LaMa model support an intelligent subtitle region selection feature similar to STTN? From my testing, the automatic subtitle detection (including STTN's automatic detection) is not nearly as accurate as when I manually select the region myself.
In my case, the videos do not contain subtitles—they only contain static text overlays. I have found that LaMa actually removes the text very well, but because it fails to detect the entire text region, some parts remain untouched. Therefore, I would really appreciate it if you could add an option that allows users to manually select the region to be processed when using LaMa, just like with STTN. I believe this would greatly improve the results for videos with fixed text overlays.
Steps to reproduce
1
Version
VideoSubtitleRemoverPro v3.18.1
Surface
GUI
Operating system
Windows 11
GPU
NVIDIA (CUDA)
Support bundle
No response
Media summary
No response
Additional context
No response