v0.6.0

this release comes with a nice quality improvement of the mosaic restoration model:

• Mosaic restoration: Removes mosaics more consistently, especially on lower-quality videos
• Mosaic restoration: Slightly fewer artifacts on certain scenes due to training on a bigger dataset
• CLI: create output directory if it doesn't exist instead of crashing
• Removed bj-pov finetuned restoration model, no improvement compared to new generic model
• Mosaic restoration dataset creation script: more accuracy for the NSFW detection model

note: the model weights lada_mosaic_restoration_model_generic_v1.2beta2.pth attached to v0.5.1-beta3 and lada_mosaic_restoration_model_generic_v1.2.pth linked to this release are the same files, just renamed.

Full Changelog: v0.5.0...v0.6.0

v0.5.1 beta3

finished retraining run started in v0.5.1 beta2. compared to beta2 there should be a slight improvement in restoring mosaics from very low-quality sources which have been rescaled/reencoded multiple times.

this version of the model weight will probably land as is in v0.5.1 unless any regressions show up.

v0.5.1 beta2

this beta comes with an updated model weights for the mosaic restoration model.
it should

1. improve restoration on videos where it previously didn't remove mosaics even though they were detected #23. issue is still present but shows up less often. given these results, there should be some potential left to improve on this a bit more i hope
2. also, slightly less artifacts in certain scenes as i doubled the size of the dataset with a bit more diverse source material

please test it on your sources and let me know if you see any regressions when using these model weights and up-to-date main branch.
download the file and move it to the model_weights directory. then specify this path when running the cli

lada-cli --input mosaic-file.mp4 --output restored-file.mp4  --mosaic-restoration-model-path model_weights/lada_mosaic_restoration_model_generic_v1.2beta.pth

v0.5.1 beta

update of nsfw detection model, now with noticeably less false-positives [nsfw detections by the model which are not actually nsfw]

Full Changelog: v0.5.0...v0.5.1-beta

v0.5.0

Compared to latest release v0.4.2 the only user-visible change this time is the inclusion of DeepMosaics pretrained restoration model in the flatpak release.

Otherwise, focused on improvements of dataset creation and reducing dependencies to make it easier to build and train new or enhance the existing mosaic restoration model. Besides refactoring there are new supplementary models (watermark (text/logo) and nudenet nsfw detection) integrated into the mosaic restoration dataset creation script. Also a slightly improved update to the NSFW detection/segmentation model.

Besides that, Lada is now also available as a docker image on DockerHub (CLI only)

Full Changelog: v0.4.2...v0.5.0

v5.0.0 beta4

Compared to beta3 we've got a pretty noticeable quality improvement of the watermark detection model.

This version of the model it is build on a new custom dataset and including some fixes in dataset creation (text/logo watermark generation).
It should be good enough now to use for filtering the mosaic restoration dataset.

Full Changelog: v0.5.0-beta3...v0.5.0-beta4

v5.0.0 beta3

Compared to beta2 there is only an incremental update of the NSFW detection model. This time, included more false-positives detected by comparing with NudeNet predictions on the same samples.

Evaluated the previously introduced watermark detection model on a bigger dataset: It is detecting a good amount of watermarks but produces many false-postives so it would eliminate a lot of valid scenes when used to create the mosaic restoration dataset.

Full Changelog: v0.5.0-beta2...v0.5.0-beta3

v5.0.0 beta2

Focus of this release are improvements to the mosaic restoration dataset creation script.
There are no changes for GUI or CLI users this time

This release comes with two additional integrated models to

1. watermarks/text/logos detection: used for filtering out obstructed scenes we shouldn't apply mosaics on
2. NudeNet NSFW detection model: used to check for false positives of Ladas primary NSFW segmentation model.

Also

• Re-trained existing NSFW model on more data which yields a bit more accurate results on my validation dataset
• faster mosaic restoration dataset creation script

Beta still as there probably needs some some tuning to find the ideal filter parameters, but can already be tested as is. You can enable them on the CLI (check out --help and/or read the code)

Note: In beta2 watermark detection model is slightly bigger but also slightly more accurate than the one from beta1

Full Changelog: v0.5.0-beta...v0.5.0-beta2

v0.5.0 beta

Focus of this release are improvements to the dataset creation scripts.
There are no changes for GUI or CLI users this time

This release comes with two additional integrated models to

1. filter out scenes obstructed by watermarks/text/logos and
2. an additional NSFW detection model to reduce number of false positives by Ladas own NSFW segmentation model.

Beta for now as there probably needs some some tuning to find the ideal filter parameters, but can already be tested as is. You can enable them on the CLI (check out --help and/or read the code)

Full Changelog: v0.4.2...v0.5.0-beta

v0.4.2

Changes

• GUI: keep current playhead position after toggling preview or doing config adjustments
• GUI: fix preview/passthrough keyboard shortcut not working
• GUI: attempt to fix stuck playback after seeking very long files

Full Changelog: v0.4.1...v0.4.2