v0.11.3

LTX is about 2.7x faster than 0.11.2 with the new default model, and 1.75x faster even if you stay on High Quality model (old default).

Support on Buy Me a Coffee or Unifans. More info
中国用户请通过闲鱼支持。其他渠道都是骗子。

Thanks to all the supporters for making it possible to train two new distilled models so quickly!

感谢所有支持者，让我们得以如此迅速地训练出两个新的蒸馏模型！

New default: Standard Quality model

Two new LTX models. Standard Quality is the new default and the fastest. Quality looks very close to me, do A/B tests and tell me if you see a difference.

Model	Speed vs Standard
Standard Quality	1x	new, default
Good Quality	~1.5x slower	new
High Quality	~2.5x slower	old default (it's still 1.75x faster compared to v0.11.2)
Top Quality	~5x slower	old Best

Speeds are for long videos. On short videos the gap is smaller.

How much faster

Same 40 s 1080p video on RTX 5090, release builds:

	0.11.2 (High)	0.11.3 (High)	0.11.3 (Standard, new default)
Whole job	186 s	106 s	68 s
Speed-up		1.75x	2.7x

Most of it comes from step 2 and step 4, which every model uses:

GPU	Step 2, per window	Step 4, per window	Step 4, Sharper large mosaics
RTX 5090	0.44 → 0.36 s	1.17 → 0.73 s	2.98 → 1.64 s
RTX 4090	0.64 → 0.49 s	1.84 → 1.07 s	4.35 → 2.45 s
RTX 5080	0.89 → 0.73 s	2.66 → 1.50 s	6.31 → 4.65 s
RTX 3090	1.37 → 1.19 s	3.72 → 2.26 s	9.10 → 5.09 s
RTX 4060 8 GB	3.04 → 2.48 s	8.83 → 6.45 s	-

Step 4 also writes the video while the next window decodes, on the 5090 the whole step went from 41 s to 19 s on that video.

RTX 50xx: Standard and Good also come in Fast mode (NVFP4), like High and Top. Fast mode is about 20% faster per step and uses about 5 GB less VRAM, so it helps most on long videos.

Other changes

• progress bar no longer jumps back and time left is more accurate
• asks before stopping, and offers to minimize when you close the window during processing
• LTX fixes for 6-8GB GPUs and long videos
• VC-1 videos on NVIDIA no longer drop frames
• videos with broken metadata text open again
• optimize end of phase 1 for LTX (could appear as stuck at 99%)
• streaming fixes
• fixed a freeze when stopping or cutting videos that are decoded on the CPU (AMD and some formats)
• Windows: activation and startup fixes, right font for Chinese text in the log
• Linux: ZFS support
• Optimize release size (linux only)
• Optimize model load time (matters most on shorter clips where loading could take most of the total processing time)

Updating

• The first run after updating is slower, it prepares the GPU once per LTX model variant.
• You can copy the LTX models from your old model_weights folder to skip the big downloads. Standard is a new model, so it downloads once.
• CLI: ltx is Standard now, ltx-4step is Good, ltx-8step is High (old default), ltx-undistilled is Top. Saved settings keep the model you picked.

v0.11.2

Read previous release notes for new model.

Support on Buy Me a Coffee or Unifans. License codes from anywhere else are fake. More info

中国用户请通过闲鱼支持。其他渠道都是骗子。

Remember that first run is slower on new version.
You can copy ltx model from model_weights directory to avoid large downloads.

Changes

• faster LTX by around 13% on long videos (longer than 1 minute) - applies to all GPU.
  Its exposed as new experimental flag that is enabled by default on Windows. It changes a bit output but I didn't notice worse quality.
  Linux has to enable it explicitly in GUI or CLI.
  You can run 1 min clip and do A/B tests to see if you spot a quality degradation. Feedback thread
• many fixes for Windows to improve startup/runtime issues and better diagnostic
• fixes for 8GB VRAM and low RAM configurations
• warning when Jasna runs on HDD (its slow, prefer SSD)
• updated attention kernels (3xxx series GPUs can see a very minor speed up, especially 3090)

v0.11.1

New model: LTX

LTX restores much more detail than BasicVSR++. It is slower and works on NVIDIA only (for now).

LTX feedback thread
Samples on Discord

• Trial mode is free, so you can check how fast it runs on your GPU. What LTX needs
• Requirements: NVIDIA RTX 30/40/50, 8 GB VRAM, 16 GB RAM (LTX uses about 10 GB). RTX 20 should work, but I only tested 2080 Ti.

LTX is for supporters

• Current supporters get LTX for free but if you feel like buying me another coffee - THANKS.
  If you didn't get the new key please write on platform you bought on.
• Support with $25 to get LTX, UNet 4x and SD 1.5 ($15 gives UNet 4x and SD 1.5).
• Your support pays for GPU time to train new models and my coffee supply.

Support on Buy Me a Coffee or Unifans. License codes from anywhere else are fake. More info

中国用户请通过闲鱼支持。其他渠道都是骗子。

No AMD build in this release. It should come with the next one, hopefully with LTX support.

How to use Jasna
Pick proper OS/GPU vendor, download all parts and unzip. You can copy model_weights folder from previous version to avoid engine recompilation.

Other changes:

• Korean and Thai translations
• post-export action can be changed during the queue, countdown before shutdown - #366
• audio/video fixes for HEVC, WMV, AAC and dropped audio - #367 #369 #390 #407 #417
• faster GUI with large queues - #355

This is big release so please report issues following the bug template - otherwise they will be just closed.

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

⚠ The only places to get a genuine key for the paid models are Buy Me a Coffee and Unifans.

⚠ 付费模型的正版密钥只能通过 Buy Me a Coffee 或 Unifans 获取。

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

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-v6and rfdetr-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-directory and 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:

• Fix encoder issue/crash/freeze for some videos #230
• Scale aware segment preview #229
• Bundle libnvJitLink for Linux builds

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