Kruk2/jasna

What it is — Jasna is a free desktop program that finds mosaic-blurred regions in videos and images and restores them, leaving everything else untouched. It has a graphical app and a command line, and does its work on your GPU rather than the CPU.

What it does

• Detects mosaics in video and restores them without changing the rest of the picture.
• Lets you restore only selected parts via a Segment Editor with a preview.
• Can play a video with mosaics restored live, without saving a file.
• Handles VR180 (side-by-side) videos and, for supporters, still images.

Two restoration models

• BasicVSR++ — fast, free for everyone.
• LTX — slower, adds more detail, but only for supporters and only on NVIDIA GPUs.

Hardware requirements

• 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.
• Alternatively AMD Radeon RX 7000/9000, marked experimental.
• Your project folder path must contain no Chinese, Japanese, accented, or other non-English characters.
• LTX specifically needs an NVIDIA RTX GPU with at least 8 GB VRAM and 16 GB RAM.
• Use an SSD, not a hard disk, or startup and model loading will be slow.

Quick start

1. Download the release for your system and GPU.
2. Unpack it into a path with only English characters.
3. Launch (jasna.exe on Windows, jasna on Linux NVIDIA, run_jasna_amd.sh on Linux AMD).
4. Add a video, pick settings, and press the play (▶) button.

The first job is slow: on NVIDIA, Jasna spends roughly 15–60 minutes preparing models for your exact GPU, once. When updating to a new version, copy the model_weights folder over to avoid repeating this. If you run out of GPU memory, lower Max Clip Size (e.g., from 90 to 60).

How it works internally Almost everything runs on the GPU, in 7 steps:

1. Decode the video on the GPU (NVDEC on NVIDIA, AMF on AMD).
2. Detect mosaics per frame (default model rfdetr-v6).
3. Track each mosaic over time, grouping frames into clips (up to Max Clip Size), with clips cut at scene changes.
4. Restore each clip as a 256×256 picture using BasicVSR++, with slight overlap so no visible jumps appear.
5. Optional secondary restoration (UNet 4x, RTX Super Res, or Topaz TVAI) makes the result bigger and sharper.
6. Blend the restored area back into the original frame — only the mosaic region changes.
7. Encode with the GPU encoder; audio is copied unless the output format can't hold it.

LTX takes a different, slower path: 4 passes (detect, prepare, restore with a diffusion model, save), with quality choices from Standard to Top (about 5× slower than Standard).

Command line — Can do most of what the app does (but not the Video Player, LTX seed previews, mask suggestions, or interactive image mode). Examples:

jasna --input input.mp4 --output output.mkv
jasna --input photo.png --output restored.png
jasna --input input_folder --output output_folder

Use jasna --help for all options.

Supporter models (paid) — Jasna itself is free. Supporting the project pays for training, mainly GPU rental. Supporters get a License code:

• $15+ total: UNet 4x (sharpens restored areas) and SD 1.5 (experimental model for still images).
• $25: the above plus LTX (NVIDIA only), including future LTX updates.
• Those who gave $15+ on or before 14 October 2026 get LTX for free.

License codes are sold only via Buy Me a Coffee, Unifans, or (for China) the seller linked in the Chinese README. Anyone else selling Jasna, codes, or modified files is flagged as a scam. One code works on 3 PCs or cloud workspaces; after activation, Jasna works offline. Old codes from version 0.10 no longer work.

Benchmarks — On an RTX 5090 (Linux), restoring a full 3:22 video with BasicVSR++: 720p H.264 in 0:32, 1080p in 0:34, 2160p in 1:03. That's roughly 3.4–4.7× faster than the comparable tool Lada, and GPU memory usage at 1080p dropped to about 4.8 GB typical / 5.6 GB peak in v0.9.1.

Licensing — The public application source is AGPL-3.0. Models and bundled dependencies keep their own terms.

Acknowledgments — Jasna is inspired by, and partly based on, the open-source tool Lada; the mosaic_restoration_1.2 model it uses was trained by Lada's author. A VR detection model was contributed by ZeLeFans.