The only places to get a genuine License code for the supporter models are Buy Me a Coffee, Unifans, and for China the seller in the Chinese README.

支持者模型的正版许可证码只能通过 Buy Me a Coffee、Unifans 或 闲鱼 获取。

English | 日本語 | 中文

Jasna

Jasna restores mosaics in videos and images. It is free, has a simple app and a command line, and runs on your GPU.

Contents

Start here

• What Jasna does
• What you need
• Quick start
• Help and community
• Supporting the project (supporter models)

Advanced

• How Jasna works
• Command line
• All guides
• Benchmarks
• Licensing
• Acknowledgments

What Jasna does

• Finds mosaics in a video and restores them. The rest of the picture is not touched.
• Two restoration models: fast BasicVSR++ (free) and slow but more detailed LTX (for supporters, NVIDIA or supported AMD GPUs; Windows AMD LTX is in beta).
• Restores only the parts you choose, with a preview, in the Segment Editor.
• Works with VR180 videos and still images (SD 1.5, for supporters).
• Lets you watch a video restored live in the Video Player, without making a file.

What you need

• NVIDIA GTX 16-series / RTX 20-series or newer, with a recent driver (610+ on Windows, 580+ on Linux). GTX 10-series and older don't work.
• Or AMD Radeon RX 7000/9000 (experimental).
• A folder on a local disk, with no Chinese, Japanese, accented or other non-English characters in its path.

More details

- NVIDIA: your GPU needs compute capability 7.5 or higher in NVIDIA's [GPU table](https://developer.nvidia.com/cuda/gpus). - AMD: a GPU supported by ROCm (Radeon RX 7000/9000, Radeon 780M/760M/740M, Ryzen AI). On Windows, AMD Adrenalin driver 26.8.1 or newer. Secondary restoration is NVIDIA only. LTX supports Radeon RX 7000/9000 and Radeon AI PRO R9700 on Linux. Windows AMD LTX is in beta and still needs GPU testing. - Disk format: NTFS on Windows; ext4, Btrfs, XFS or ZFS on Linux. Network disks are not supported. Use an SSD; on a hard disk (HDD) Jasna starts and loads models slowly. - LTX needs an NVIDIA RTX GPU or a supported AMD GPU, at least 8 GB VRAM and 16 GB RAM. See [What LTX needs](https://github.com/Kruk2/jasna/blob/HEAD/docs/en/models.md#what-ltx-needs).

Quick start

1. Download the release for your system and GPU.
2. Unpack it into a folder with no Chinese, Japanese, accented or other non-English characters in the path.

3. Start the app: - Windows: double-click jasna.exe. - Linux NVIDIA: run the jasna file. - Linux AMD: first install AMD's AMF video runtime once. We tested AMF 26.20:

   bash curl -fLO https://repo.radeon.com/amf/26.20/amf_installer_26.20.zip unzip amf_installer_26.20.zip sudo bash amf_installer_26.20.sh

   Accept AMD's license when prompted, then run run_jasna_amd.sh. If using Distrobox, install AMF inside the same container as Jasna. 4. Add a video, choose settings and click the ▶ button at the bottom.

Your first job is slow. On NVIDIA, Jasna first prepares the models for your exact GPU. This takes about 15–60 minutes, only once. Close other apps and don't use the PC during this time. When you update Jasna, copy the model_weights folder to the new version, so you don't wait again.

LTX update in 0.12: models copied from 0.11.3 need downloading again, including on NVIDIA. Press ▶ and accept the model update. Your existing License code and shared VAE files still work. For manual model downloads, use branch 0.12.0 for Jasna 0.12, or main for 0.11.3 and older.

Every setting has a tooltip: move the mouse over the ⓘ icon. Out of GPU memory (VRAM)? Lower Max Clip Size, for example from 90 to 60. See Tuning.

Help and community

• App guide: the queue, presets, output names, the Video Player and activation.
• SLS Discord: examples, help and settings. Jasna has a new channel; older posts stay in the old channel (archived). Please don't be too weird.
• Problem? Click Logs at the bottom of the app, then Export to save a log file. Attach it when you ask for help.

Supporting the project

Jasna is free. Your support pays for training new models, mostly GPU rental. Supporters get a License code for the supporter models:

Support	Models
$15 USD or more in total	UNet 4x: makes restored areas sharper. SD 1.5: experimental model for still images.
$25 USD	UNet 4x and SD 1.5, plus LTX: much slower than BasicVSR++, but more detail. NVIDIA or supported AMD GPUs; Windows AMD LTX is in beta. Includes future updates of this LTX model.

Supporters who gave $15 USD or more on or before 14 October 2026 get LTX for free.

Before you choose LTX:

• Look at these LTX samples: 1, 2, 3, 4, 5, 6, 7, 8. Explore the new channel and the old channel (archived) on SLS Discord for more samples.
• Run LTX in Trial mode (speed test) on your PC. It is free and shows if your GPU can run LTX and how long it takes. The video is not restored. See What LTX needs.

More examples: UNet 4x, SD 1.5, more SD 1.5.

Where to support

License codes come only from here:

• Buy Me a Coffee (crypto works too)
• Unifans
• For China: see the Chinese README

Anyone else selling Jasna, codes or modified files is a scam: their keys don't work, and their files can run hidden code on your PC. Download Jasna only from GitHub.

How to activate

1. Support on one of the pages above. You get a License code.
2. In Jasna, open Activate Jasna, enter the License code and press Activate online.
3. No internet on that PC, or it fails? Choose Manual, see Activate supporter models.

One License code works on 3 PCs or cloud workspaces. After activation, Jasna works offline. Rented GPUs: see Cloud GPUs.

After activation, opening Activate Jasna checks online and shows how many device slots are used.

What counts as a new PC: on Windows, a new motherboard; on Linux, a fresh Linux install. Each uses one of your 3 slots. Reinstalling Windows or changing the GPU, disk or RAM does not: just activate again with the same License code. A few Windows PCs have no real motherboard ID; on those, a new system disk also counts as a new PC.

Have an email and key from Jasna 0.10? They no longer work. Get a new License code from where you bought it.

Advanced

How Jasna works

Almost everything runs on the GPU, from reading the video to saving it.

1. Decode the video on the GPU (NVDEC on NVIDIA, AMF on AMD).
2. Detect mosaics in each frame (default model rfdetr-v6).
3. Track each mosaic over time and group its frames into clips (up to Max Clip Size frames). Clips end at scene cuts, so two different shots are never mixed.
4. Restore each clip as a 256x256 picture with BasicVSR++. Clips overlap a little, so there are no visible jumps between them.
5. Optional secondary restoration (UNet 4x, RTX Super Res or Topaz TVAI) makes the restored picture bigger and sharper.
6. Blend the restored picture back into the original frame. Only the mosaic area changes.
7. Encode with the GPU encoder (NVENC on NVIDIA, AMF on AMD). Audio is copied (encoded again only if the output format can't hold it).

The first job is slow because the models for steps 2 and 4 (and UNet 4x, if used) are prepared for your exact GPU (NVIDIA only). The result is saved in model_weights and used again later. You can copy it to a newer Jasna version.

LTX works differently. It goes over the video in 4 passes: find mosaics, prepare, restore with a diffusion model, save. It is slower but can add more real detail. Its quality choices: Standard (default, fastest), Good (about 1.5x slower), High (about 2.5x slower) and Top (about 5x slower). See Choosing models.

LTX jobs can pause and continue, including range jobs and auto split. See Pause and continue.

After an update, Jasna may ask to download your chosen LTX model again. Your existing License code still works.

Command line

The command line can do most of what the app does (not the Video Player, LTX seed previews, mask suggestions or interactive image mode).

# One video
jasna --input input.mp4 --output output.mkv

# One image
jasna --input photo.png --output restored.png

# A whole folder
jasna --input input_folder --output output_folder

jasna --help lists all options. See the CLI reference.

All guides

• Using the GUI: queue, presets, output names, Video Player, activation.
• Choosing models: detection models, BasicVSR++ or LTX, secondary restoration, SD 1.5 for images.
• Restoring only parts of a video: Segment Editor, mosaic scan, LTX only for big mosaics.
• VR180 videos: how side-by-side VR is handled.
• Tuning VRAM and GPU usage: clip size, overlap, what to do when VRAM runs out.
• Advanced processing: denoise, 60→30 FPS, color LUT, sharpening, encoder quality, actions after the queue.
• Streaming: watch in the browser or in Stash.
• Cloud GPUs: Jasna on a rented GPU (RunPod, Vast).
• CLI reference: all command line options.
• Running from source: for developers. From source you can use the free BasicVSR++; activation and supporter models need the official release.

Benchmarks

Time to restore a whole 3:22 video with BasicVSR++ on an RTX 5090 (Linux). Lower is better.

Input	Lada 0.11.0	Jasna v0.9.1
720p H.264	01:47	00:32 (3.4x faster)
1080p H.264	02:02	00:34 (3.6x faster)
2160p H.264	04:56	01:03 (4.7x faster)

GPU memory, older versions and how it was measured

GPU memory used (typical / peak, in GB). `—` means not measured. Lada uses less GPU memory than Jasna. | Input | Lada | v0.4.1 | v0.5.0 | v0.9.0 | **v0.9.1** | | --- | ---: | ---: | ---: | ---: | ---: | | 720p H.264 8-bit | 1.8/3.1 | 17.6/19.9 | 8.6/9.0 | 8.6/9.1 | **4.2/4.6** | | 1080p H.264 8-bit | 2.0/3.3 | 18.9/21.4 | 9.5/10.2 | 9.3/10.1 | **4.8/5.6** | | 2160p H.264 8-bit | 2.6/4.1 | 26.0/30.5 | 12.9/15.0 | 11.4/15.4 | **7.2/10.9** | | 8K VR HEVC 60 fps (15 s) | — | — | — | 17.1/18.3 | — | Times for older versions (v0.4.1: 01:35 / 01:46 / 03:23; v0.5.0: 00:45 / 00:47 / 01:22; v0.9.0: 00:34 / 00:39 / 01:15 for 720p / 1080p / 2160p; 8K VR: 00:34), the test settings, RAM use and every codec are in the [full benchmark report](https://github.com/Kruk2/jasna/blob/HEAD/benchmarks/2026-07-26_release_matrix.md). Older full-length videos: [legacy report](https://github.com/Kruk2/jasna/blob/HEAD/benchmarks/2026-07-26_legacy_videos.md).

Licensing

Jasna's public application source is AGPL-3.0. Models and bundled dependencies keep their own terms. See licensing and source availability, model notices and the release source correspondence.

Acknowledgments

• Lada (Codeberg): Jasna is inspired by Lada and in some places based on it. The mosaic_restoration_1.2 model used by Jasna was trained by ladaapp, the author of Lada.
• ZeLeFans: for the VR detection model and the detailed analysis of mosaic shapes and VR projection.