Releases: Kruk2/jasna
Releases · Kruk2/jasna
Release list
v0.10.0
🚀 If you want to support this project Buy me a coffee or Unifans 🚀
By supporting you will get access to two models
You can copy .engine files from previous version to avoid engine recompilation (just copy entire model_weights folder and override).
How to use Jasna
Pick proper OS/GPU vendor, download all parts and unzip.
v0.10.0
New features:
- High performance Video Player. You can click the button at the top or add multiple files to Queue and play them
- Custom command execution after video is completed. Read more
- Context menu for videos in queue (open location, copy path, requeue)
- Explicit button to output restored video to the same folder as input
- TVAI denoise filter - @devxyius #261 (it's slow, expect warnings with stall/flush retry)
- Support segment processing for AMD cards
Fixes:
- many AMD fixes - #286 #252. Thanks @RainerGevers
- h264 CQ control - fixes #282
- preserve video metadata - fixes #283
- improved frame rate retargetting (from 60fps to 30fps)
- improved streaming deadlock/crashes
- fixed video decoder for some special videos - fixes #279
- improved audio channels logic - fixes #273
Contributors
RainerGevers and devxyius
Assets 11
v0.9.1
🚀 If you want to support this project Buy me a coffee or Unifans 🚀
By supporting you will get access to two models
You can copy .engine files from previous version to avoid engine recompilation (just copy entire model_weights folder and override).
How to use Jasna
Pick proper OS/GPU vendor, download all parts and unzip.
Check also v0.9.0 release notes
v0.9.1
Improvements:
- Massive VRAM usage reduction when using compiled restoration model.
4.5GB less on average for 180 clip size and 2.8GB less on 90 clip size
Check benchmark table - Increasing clip size still increases VRAM usage but at much lower rate
- Compilation is now clip-size agnostic which means that you can change it and model won't recompile
- Allow for up to 720 clip size in GUI
- More speed optimizations (~12% faster on average)
Fixes:
- more AMD fixes
- more Windows scaling fixes (app freezes on startup)
v0.9.0
🚀 If you want to support this project Buy me a coffee or Unifans 🚀
By supporting you will get access to two models
You can copy .engine files from previous version to avoid engine recompilation (just copy entire model_weights folder and override).
Usage
https://github.com/Kruk2/jasna#quick-start
Pick proper OS/GPU vendor, download all parts and unzip.
Changelog
New models
- No new restoration models :) but there are some previews on Discord
- Added two new detection models:
rfdetr-v6andrfdetr-v6-large. Both have similar accuracy but rfdetr-v6 is 2x faster compared to rfdetr-v5. V6-large can handle 4K content better. New models have higher default detection threshold so if you used a custom one then you should review it. v6-large is a separate download (check models docs) - Added detection model dedicated for VR:
rfdetr-vr-v1. Zelefans VR model is no longer included by default - Improved mosaic detection logic. Added two new settings to control how miss-detection and false mosaic detection are handled. Jasna will now by default try to fill missed detection by interpolating the mask. This should reduce mosaic flickering issues in challenging scenes. Read more
New features
- Performance optimizations. Around 15% faster for 2160p and 25% for 8k videos and similar VRAM (tested on Linux, RTX 5090, i9-13900k). Benchmarks
- implemented Scene Cut Detection (default ON). This helps in cases where there was hard scene cut yet mosaic placement in new scene overlaps with old scene. Read more
- improve mosaic scan times (up to 2-3x faster on my 5090) in Segment Editor window. Other GPUs might benefit less since only RTX 5090/5080 have two decoder chips.
- CAS sharpening filter (might improve final image especially for hairs and other fine details). Fixes #246 Read more
- add ability to zoom/pan in the Segment Editor window
- improve default encoding settings (lower file size, same quality). To benefit from it you have to reset your preset. Fixes #235
- many AMD fixes for windows and linux. Fixes #240
- fragmented MP4 (default OFF) so you can play video while its processing. Fixes #244
- Added Gnomonic projection for VR mosaic - I found that for some videos it produces even better result (not very common though).
- Added back
--working-directoryand exposed it in GUI. Fixes #250
Fixes
- make the startup window a bit larger
- fix hardware decoding for AV1 sources. Fixes #237
- remove adding metadata for VR videos. Fixes #231
- HDPI. Fixes #241
- fix retartgeting FPS for videos with incorrect metadata. Fixes #248
- improve VR auto-detection settings
- handle ffmpeg crash in streaming mode
- handle incorrect DTS/PTS in input video (regression). Fixes #248
- fix STOP button behavior in GUI
v0.8.1
🚀 If you want to support this project Buy me a coffee or Unifans 🚀
You can copy .engine files from v0.6/v0.7/v0.8 versions to avoid engine recompilation (just copy entire model_weights folder and override).
It requires 610 Nvidia driver on Windows. Linux stays at 580 (for now).
Usage
https://github.com/Kruk2/jasna#usage
Pick proper OS/GPU vendor, download and unzip.
Changelog: v0.8.0-> v0.8.1
Release notes for v0.8.0
Features:
- Experimental AMD builds for linux/windows. Probably doesn't work, I don't have a way to test it but if you have AMD feel free to give it a go.
Fixes:
v0.8.0
🚀 If you want to support this project Buy me a coffee or Unifans 🚀
You can copy .engine files from v0.6/v0.7 versions to avoid engine recompilation.
Since this release is big, expect some bugs. If you encounter any issues please report them (and follow the bug template).
It requires 610 Nvidia driver on Windows. Linux stays at 580 (for now).
No new model in this release.
Usage
https://github.com/Kruk2/jasna#usage
Changelog: v0.7.x -> v0.8.0
Features:
- Segment processing. Check Readme for more details.
- Segment processing also allows for:
- fast pre-processing of video to find where mosaics are located
- previewing restoration result
- suggesting improved mask if you think the detection model was wrong. This might be used later for training a better model.
- VR support with model from zelefans that might be better than rfdetr-v5 in some cases. Thanks Ganonmaster for first PR. Keep in mind that restoration model was not trained for VR so quality might vary. You can use the preview tool mentioned above.
- AV1 and H264 encoding
- Fallback to slower, software decoding for unsupported input codecs
- Removed the need of mkvtool, this also removed working directory option as the full output is written in place.
- Add option to re-target FPS for faster processing. You can now process 60FPS videos 2x faster by converting them on the fly to 30FPS
Fixes:
- Handle corrupted input videos. Jasna will try to process corrupted videos (to some extend) instead of failing hard on first bad frame.
- Linux GUI fixes and general UI/UX improvements
- Improved colorspaces support #75
- FP16 fix for unet-4x #201
- Handle old videos with sample aspect ratio (anamorphic) #125
v0.7.2
🚀 If you want to support this project Buy me a coffee or Unifans 🚀
You can copy .engine files from v0.6/v0.7 versions to avoid engine recompilation.
Usage
https://github.com/Kruk2/jasna#usage
updated linux build, if you downloaded previous one and it fails with libavformat.so.62 or similar, download it again.
Check also previous release notes
Changelog: v0.7.1 -> v0.7.2
Features:
- Support for yolo detection models in SD 1.5 image restoration
- You can now specify action once restoration is done (shutdown computer/custom command). More details
- Cleanup README
- Added option to specify custom file patterns when processing directory in CLI. Check Quick Start section.
- Linux now bundles ffmpeg8 and mkvmerge so no need to extra installation.
Fixes:
- Windows GUI randomly bugging (randomly fails to restore and/or fails system check)
- Optimized GUI startup time
v0.7.1
🚀 If you want to support this project Buy me a coffee or Unifans 🚀
You can copy .engine files from v0.6 version to avoid engine recompilation.
Usage
https://github.com/Kruk2/jasna#usage
Changelog: v0.7.0 -> v0.7.1
- Fixed npp cuda loading on windows & bundle mkvmerge & ffmpeg
- Fix image restoration model download from GUI (if you downloaded sd 1.5 model on 0.7.0 you have to download it again)
- Add an interactive window to image restoration (accessible in "Image restoration" section). You can add multiple images, and see/probe different restorations:
Changelog: v0.6.x -> v0.7.0
Features:
I'm publishing two models, currently available only for people that are supporting me on buymeacoffee/unifans as a way to say thank you.
More details
- Added unet-4x model for secondary restoration. I think it's less grainy than RTX Super Res while also quite fast. Comparable to TVAI iris-2, if you notice any weird artifacts in restored videos let me know. Details on vram/time/example restorations
- Added Diffusion model to restore STILL images. This is old checkpoint that I trained some time ago. It's not perfect but when it works it looks very good. Probably struggles with uncommon scenes/small mosaic. I can't post NSFW here but if you have access to SLS discord you can see few previews. More details
Fixes:
- RTX super resolution should now work on linux
- Fix TVAI integration for some corner cases
v0.6.2
🚀 If you want to support this project buy me a coffee 🚀
You can copy .engine files from v0.6.1 version.
Added linux build as well.
Changelog: v0.6.1 → v0.6.2
Fixes:
v0.6.1
🚀 If you want to support this project buy me a coffee 🚀
Don't copy .engine files from previous version. Tensorrt update requires engines to be rebuild.
Changelog: v0.5.x → v0.6.1
Features:
- [experimental] LUT color corrections via cube files. Fixes #142 - you can find there example cube files
- vram and performance improvements (thanks to sh202603 for finding O(n^2) regression in dense conv).
Your results might be different, depends on GPU and restored area resolution.
| Video | Before (time / VRAM) | After (time / VRAM) | Speed Δ | VRAM Δ |
|---|---|---|---|---|
| file-1080p-A | 4m 50s | 4m 15s | -12.1% | -14.8% |
| file-1080p-B | 57s | 54s | -5.3% | ~-0.6% |
| file2-1080p | 32s | 31s | -3.1% | -12.7% |
| 4k | 1m 45s | 52s | -50.5% | -19% |
- bt 601 colorpsace support (color space used in old videos).
- streaming support either via built-in app that you can open in browser or via my stash fork.
Stash is prefered but you can start streaming mode in your browser viajasna-cli --stream. Specify other settings if you prefer.
STASH
Fixes:
- model_weights are not read from current working directory. fixes #161 #133
- improved HAGS detection (yes you still have to disable hardware accelerated scheduling on windows) & added nvidia driver check & path check.
- relax NVIDIA driver requirement to 580+ fixes #117
- nvidia-smi checks improvements #124
v0.6.0-alpha5
🚀 If you want to support this project buy me a coffee 🚀
Remember to copy model_weights folder from previous version to new to avoid engines compilation.
alpha5
With dedication to friends at Sehuatang, especially those who don't read instructions.
- you can't skip system check if you fail it
- improved HAGS detection & added nvidia driver check & path check
致献给 Sehuatang 的朋友们，尤其是那些不看说明的人：
- 如果系统检查未通过，就不能跳过
- 改进了 HAGS 检测，并新增了 NVIDIA 驱动检查和路径检查
alpha4
- fixed CLI output path
alpha3
- streaming improvements
- integration with stash via my fork so you can watch any stash video processed by Jasna (on-the fly)
If you didn't use stash yet I highly recommend it for organizing/rating/managing your JAV collection.
Now you can also decensor any video on the fly in stash - without the need to process entire file beforehand.