项目的 README,来自 github.com/Kruk2/jasna。合并为一个 markdown 文件:README.md。
The only places to get a genuine License code for the supporter models are Ko-fi, Unifans, and for China the seller in the Chinese README.
Jasna restores mosaics in videos and images. It is free, has a simple app and a command line, and runs on your GPU.
Start here
Advanced
| AMD GPU | Jasna | LTX |
|---|---|---|
| Radeon RX 9060/9070, Radeon AI PRO R9700 | yes | yes (Windows: beta) |
| Radeon RX 7600–7900, Radeon PRO W7000 | yes | yes (Windows: beta) |
| Radeon 740M/760M/780M/880M/890M, Ryzen AI Max (8050S/8060S) | yes, untested | yes, untested |
| Radeon RX 6000, Radeon PRO W6000 | yes, untested | no |
| Radeon 610M/660M/680M/820M/840M/860M | yes, untested | no |
| Radeon RX 5000, Radeon PRO W5000, RX Vega, RX 500 and older | no | no |
Windows: double-click jasna.exe. On AMD, first install
AMD's driver for ROCm 10.1
(26.10.41.05, download .exe)
and reboot. You don't need to install ROCm.
Linux NVIDIA: run the jasna file.
Linux AMD: first install AMD's AMF video runtime once. We tested AMF 26.20:
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.
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 may
still ask to download its model again after an update. On AMD, only UNet 4x
needs this preparation, on its first use.
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.
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 (Ko-fi: 14 EUR) or more in total | UNet 4x: makes restored areas sharper. SD 1.5: experimental model for still images. |
| $25 USD (Ko-fi: 23 EUR) | UNet 4x and SD 1.5, plus LTX: much slower than BasicVSR++, but more detail. 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:
License codes come only from here:
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.
One License code works on 3 devices at the same time. On your own PC with a TPM 2.0 security chip (most Windows 11 PCs), Jasna then works offline. Elsewhere it needs internet while it uses supporter models. Details and help: Activate supporter models. Rented GPU: see Cloud GPUs.
Jasna finds each mosaic, follows it from frame to frame, restores only that area and puts it back. The rest of the picture is not touched. Almost everything runs on the GPU. LTX is slower but can add more real detail; see Choosing models.
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.
On an RTX 5090 (Linux), Jasna 0.9.1 restored a 3:22 1080p video with BasicVSR++ in 34 seconds. Test settings, GPU memory and older versions are in the full benchmark report; newer results are in benchmarks.
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.
mosaic_restoration_1.2 model
used by Jasna was trained by ladaapp, the author of Lada.