English | 日本語 | 中文

Supporters get a key that unlocks the extra models I train — the unet-4x secondary upscaler and the experimental SD 1.5 image restoration model. Details in Supporting the project.

JAV model restoration tool inspired (and in some places based on) by Lada.
Restoration model (mosaic_restoration_1.2) used in Jasna was trained by ladaapp (the lada author).

Features new mosaic detection model & super fast GPU only pipeline & TVAI support & simple GUI.
Check benchmarks and usage below.

• Improved mosaic detection model.
• Temporal overlap which reduces flickering.
• Secondary restoration model (Topaz TVAI and RTX Super Res). Improves quality and sharpnes
• Experimental SD 1.5 image restoration for still photos — sometimes produces really good results (supporter model).
• GPU only processing. Intial tests show that it can be 2x faster. Raw processing for places without mosaic is ~250fps on RTX 5090
• Accurate color conversions on gpu (input matches output and no banding).
• Only modern Nvidia gpu is supported.
• TensorRT support.

• proper VR support
• TVAI and SeedVR
• Proper stream that can be played in Stash (and maybe others?)
• improve performance (this version is very simple)
• improve VRAM usage

Jasna automatically manages VRAM. When your GPU runs low on VRAM, frames waiting in the processing queue are temporarily moved to system RAM and moved back when needed. This happens in the background and requires no configuration.

If you still run into out-of-memory errors, reduce the clip size (e.g. from 180 to 60) or disable model compilation — both lower peak VRAM significantly (see table below).

RTX 5090 + i9 13900k

Support pays for training the extra models — mainly renting bigger GPUs and the compute time to train on larger datasets. Supporters get a key that unlocks the models I train this way:

• unet-4x secondary upscaler (sharper 256→1024 restoration).
• SD 1.5 image restoration (experimental still-image model).

How to get a key:

1. Contribute $15 or more in total — across any number of contributions, at any time.
2. Email me at myprotonmailkekw@proton.me from the address/handle you used, and I'll send your key. The key is tied to that email.

Buy me a coffee handles contributions, including crypto.

Download the latest release package (Windows/Linux).

• If you downloaded the app (recommended):
  • Windows: You’re good to go — Jasna ships with everything it needs (ffmpeg, ffprobe, and mkvmerge).
  • Linux: You need ffmpeg, ffprobe (major version must be 8), and mkvmerge available on your system. Install via your package manager. MKVToolNix: downloads.

First run will be slow — TensorRT engines are compiled for your GPU. This takes 15-60 minutes.
Close all other applications (including browsers) and do not use the PC during compilation.
Engines are cached in the model_weights folder and reused on all future runs (you can copy engine files & folders to a new version).

Remember to have up to date nvidia drivers.
Tested nvidia drivers: 591.67 (but anything from 59x family should be ok and it's minimum required).
Jasna requires GPU with minimum compute capability: 7.5

In general it's recommended to pick latest rf-detr.
Lada Yolo models are available as they handle 2d animations better.

Jasna/Lada takes 256x256 pixel crop of mosaic region and restores it still in 256x256 resolution. This causes blurry results when mosaic area was bigger (close ups, 4k video etc).
To eleviate that, you can use 2nd restoration model that upscales 256x256 to 512x512 or 1024x1024 which produces cleaner results. Currently supported:

• unet-4x (supporter model). Trained on an in-domain (JAV) dataset and visually quite close to TVAI iris-2, but runs locally with no extra setup. If you run into quality issues, please open a GitHub issue. Unlock it with a supporter key — see Supporting the project.
• RTX Super-resolution (very fast, ok quality). Very fast, free, zero dependencies. In some videos it may produce a flickering effect — test on a short clip first. Place jasna in folder with english only characters.
• TVAI (best quality, slowest). Requires Topaz Video (paid, Windows only). Recommended model: iris-2.
  --tvai-args allows you to customize model and other params. Defaults to iris-2.
  Setup these as env variables for "Topaz Video":
  

For still images Jasna can use a fine-tuned Stable Diffusion 1.5 inpaint model instead of the video pipeline. It detects mosaics, inpaints each region at 512×512 and blends the result back.

For examples on restorations you can checkout SLS discord.

• CLI: jasna --input photo.png --output out.png — image inputs auto-route to the SD 1.5 model, no flag needed. The image model is selected with --image-restoration-model-name (default and only value: sd-15-jav); --restoration-model-name stays video-only. Knobs: --sd15-steps, --sd15-strength (≤ 0.7), --sd15-freeu/--no-sd15-freeu, --sd15-seed, --sd15-variants N.
• GUI: just add an image to the queue — image jobs route to the SD 1.5 model automatically. Tune it in the Image Restoration settings section.
• Folder input (CLI): --input <folder> --output <folder> processes every media file in the folder — images first, then videos — writing <name>_out<ext> into the output folder (--output must be a folder here). The GUI queue already mixes images and videos freely.
• The model is not bundled (~6.9 GB). It lives in model_weights/sd-15-jav/. Either drop the bundle there yourself, or let Jasna fetch it from huggingface.co/Kruk2/sd-15-jav — it asks before downloading (CLI prompt, or the Download model button in the GUI). The checkpoint is encrypted — you need a supporter key to use it (same key as unet-4x — see Supporting the project).
• It's experimental: results vary by scene, but on the right image it can look genuinely good. Try a few --sd15-variants (different seeds) and keep the best one.
• Expect around 7 GB of VRAM during inference (a bit more on large 4K images).

Main goal for the temporal overlap is to reduce flickering on the edges of clips.
Going above 20 doesn't bring much. The larger overlap the longer processing times but less flickering.
Pick highest clip size you can and set overlap to something between 8-20.
Some guidance from my limited testing:

• 60 clip size + temporal overlap 6
• 90 clip size + temporal overlap 8
• 180 clip size + temporal overlap 15 (needs 12 GB+ VRAM with Compile BasicVSR++ enabled, less with it disabled).

4K videos use more VRAM — a lower clip size may produce similar quality but process much faster. Experiment a bit.

Using clip size below 60 might look ok, depends on video but prefer using 60 even if it means disabling model compilation. --enable-crossfade should probably be always enabled as it reduces flickering .

The restoration model is compiled into TensorRT sub-engines.
First compilation takes 15-60 minutes — close all other applications (including browsers) and avoid using the PC.
Engines are cached and reused automatically. You can opt out from compilation at the cost of performance.

The table below shows how much VRAM the compiled engines alone occupy (not total processing VRAM):

Watch restored videos on-the-fly without having to process the entire file first.

Browser player (built-in)
Run jasna with --stream flag (CLI only for now). A browser window opens with an HLS player. Pick a video file and start watching. Seeking is supported.

jasna --stream

Stash integration
Use Jasna directly inside Stash — play any scene and it gets processed through Jasna on-the-fly. Stash launches Jasna automatically. A custom Stash fork with Jasna support is available:
👉 Stash v0.30.1-jasna

Setup:

1. Download the Stash fork from the link above.
2. Set environment variables before starting Stash:
   • JASNA_CLI_PATH — full path to jasna-cli.exe
   • JASNA_WORKING_DIR — full path to the folder where jasna-cli.exe is located
3. Start Stash and play any scene — Jasna processes it as you watch. Seeking works.

Jasna is in early development and the main goal is to improve: restoration quality, mosaic detection, speed & vram consumption (in this order). Thats why currently this project is aimed at more technical users, meaning the program accessability is lower and might be gated for some users. I do this to dedicate more time on important features and if you want to help Pull Requests are welcomed.

Install these libs via uv pip install . --no-build-isolation
To build nvidia libs below you need also VS Build Tools 2022 (c++) make sure you have cmake and ninja uv pip install cmake ninja
and cuda 13.0 in your system.

• Install ffmpeg + ffprobe and make sure they are on PATH (ffmpeg major version must be 8).
• Install mkvmerge (part of MKVToolNix): downloads.

https://codeberg.org/Kruk2/vali

https://codeberg.org/Kruk2/PyNvVideoCodec

Once two libs above are installed to your python enviorment run: uv pip install -e .[dev] in jasna repository