Who Jasna is for#
Video editors processing locally stored footage
Jasna's Segment Editor and batch CLI make it practical for editors who process large volumes of video regularly. The free BasicVSR++ tier handles most standard workflows without ongoing cost, and the GPU-accelerated pipeline finishes a typical clip in seconds rather than minutes.
Skip if:
Skip if your footage is stored on network-attached storage or in a path with non-ASCII characters; Jasna supports neither. Also skip if you need cloud-based processing or real-time collaboration on shared infrastructure.
Privacy-focused researchers avoiding cloud uploads
All processing stays on local hardware with no API calls to external services. Researchers handling sensitive video material can run the full restoration pipeline in an air-gapped environment. The AGPL-3.0 source code is auditable for anyone who needs to verify what the tool does.
Skip if:
Skip if your research environment has no qualifying NVIDIA or AMD GPU (GTX 16-series/RTX 20-series or newer, or RX 7000/9000). A CPU-only fallback is not available.
Developers building automated video processing pipelines
The CLI accepts single files or entire folders and exposes most GUI settings as flags, making Jasna scriptable for automated pipelines. Running `jasna --input input_folder --output output_folder` processes a whole directory batch. Cloud GPU support via RunPod or Vast enables scaling beyond local hardware.
Skip if:
Skip if your pipeline needs to run on non-NVIDIA hardware in production, or if you require a REST API or container-native deployment. Jasna is a desktop application, not a microservice.
The problem it solves#
Removing mosaic censorship from video footage has no free, automated path that scales to full-length content. Commercial tools like UniFab Video Enhancer charge per license or subscription and process files on vendor infrastructure, creating privacy concerns for anyone working with sensitive footage. Older open source options exist but are slower. Lada 0.11.0, a comparable tool, takes over two minutes to restore a 3:22 1080p video on hardware where Jasna finishes in 34 seconds.
The difficulty multiplies with longer videos. Each file contains thousands of frames with mosaic regions that must be detected and tracked consistently across scene cuts. Without GPU-accelerated detection and temporal tracking, manual correction at scale is not practical.
How it solves it#
Full GPU pipeline from decode to encode
Video decoding (NVDEC/AMF), mosaic detection, tracking, restoration, and output encoding all run on the GPU. On NVIDIA, the encoder uses NVENC; on AMD, AMF. Audio is copied without re-encoding unless the container format requires it, which preserves quality and reduces processing time.
BasicVSR++ and LTX model tiers
BasicVSR++ (free) restores a 3:22 1080p video in about 34 seconds on an RTX 5090. The LTX diffusion model (supporter tier, $25 USD, NVIDIA RTX only) is slower but adds more detail through a 4-pass process (detect, prepare, diffusion restore, save). LTX requires at least 8 GB VRAM and 16 GB RAM.
Segment Editor for region-specific restoration
The Segment Editor lets you pick exactly which mosaic areas get processed. Restoration runs only on the chosen regions; the rest of the frame is untouched. Supports mosaic scan to auto-suggest regions, with manual override, so you can skip false detections or add missed areas.
CLI for batch video and image processing
The command line accepts a single video, a single image, or a whole folder as input. Passing a folder path batches an entire directory in one run. Most GUI settings are available as CLI flags; `jasna --help` lists all options and the full reference is in docs/en/cli.md.
VR180 video and still image modes
Side-by-side VR180 videos are handled with projection-aware processing so both eye views are restored consistently. Still image restoration uses the SD 1.5 model (supporter tier, $15 USD or more). Both extend Jasna beyond standard flat video for users working with less common content formats.
Live Video Player preview
The Video Player mode plays back restored footage in real time without writing an output file. This is useful for previewing restoration quality and tuning settings like clip size or model choice before committing to a full encode run on a long video.
Strengths and trade-offs#
Strengths
- 3.4x to 4.7x faster than Lada on equivalent hardwareBenchmarks against Lada 0.11.0 on an RTX 5090 show Jasna v0.9.1 restoring 720p footage 3.4x faster (00:32 vs 01:47) and 4K footage 4.7x faster (01:03 vs 04:56). VRAM usage also dropped: v0.9.1 uses 4.2 to 4.6 GB peak for 720p compared to v0.4.1's 17.6 to 19.9 GB.
- Runs entirely on local hardwareAll processing stays on the user's GPU with no files sent to external servers, no API dependency, and no per-video billing. Supports NVIDIA GTX 16-series or newer (Windows/Linux) and AMD RX 7000/9000 experimentally. Cloud GPU rental via RunPod or Vast is documented for users without local GPU hardware.
- AGPL-3.0 licensed and auditableThe application source is AGPL-3.0, so anyone can inspect, modify, and run the code. Models and bundled dependencies have their own terms (see LICENSING.md and THIRD_PARTY_MODELS.md in the repo). The free tier with BasicVSR++ has no ongoing cost beyond the hardware to run it.
- Active development with rapid optimizationCreated in January 2026, Jasna had its last commit on October 18, 2026, with 2,426 stars and 187 forks. GPU memory use dropped from 17.6 to 19.9 GB peak (v0.4.1) to 4.2 to 4.6 GB (v0.9.1) for 720p footage, showing consistent performance improvement across releases.
Trade-offs
- -NVIDIA required for full feature accessThe LTX diffusion model and UNet 4x secondary restoration require an NVIDIA RTX GPU. AMD support is experimental (RX 7000/9000 series only), with no LTX or secondary restoration available on AMD. GTX 10-series and older are not supported at all. Users without qualifying NVIDIA hardware are limited to BasicVSR++ with AMD or must use cloud GPUs.
- -Supporter models require a paid license codeUNet 4x sharpening and LTX diffusion restore are available only with supporter contributions ($15 USD for UNet 4x and SD 1.5; $25 USD for all three including LTX). These are one-time contributions rather than recurring subscriptions, but the features are not freely available. One license code works on up to 3 PCs or cloud workspaces.
- -First run takes 15 to 60 minutes on NVIDIAOn NVIDIA hardware, Jasna compiles and prepares model weights optimized for the exact GPU during the first run. This takes 15 to 60 minutes. The compiled weights are cached in the model_weights folder and can be copied to a newer Jasna version, but a fresh install on a new machine triggers re-preparation.
- -Local disk and ASCII path requiredThe working folder must be on a local disk (NTFS on Windows; ext4, Btrfs, XFS, or ZFS on Linux) with no Chinese, Japanese, accented, or other non-ASCII characters anywhere in the path. Network-attached storage is not supported. This is a barrier in environments with shared or remotely mounted filesystems, or paths in non-Latin character sets.
Jasna vs alternatives#
Jasna vs UniFab Video Enhancer
Both tools detect and restore mosaic censorship in video footage. The decision comes down to cost model, data privacy, and workflow flexibility.
| Feature | Jasna | UniFab Video Enhancer |
|---|---|---|
| License | AGPL-3.0 | Proprietary |
| Cost | Free (base tier); $15-$25 for advanced models | Per-license commercial |
| Processing location | Local GPU only | Vendor infrastructure |
| GPU support | NVIDIA GTX 16+, AMD RX 7000/9000 (exp.) | NVIDIA and AMD |
| CLI | Yes | No (GUI only) |
| Batch folder processing | Yes | Limited |
| VR180 video support | Yes | No |
| Diffusion model option | LTX (supporter tier) | Proprietary AI engine |
Jasna is the better choice when you need local processing with no cloud uploads, a scriptable CLI for batch workflows, or an auditable codebase. The base tier costs nothing beyond the hardware, and the one-time supporter contribution for advanced models is a fixed cost rather than a recurring subscription.
UniFab Video Enhancer is worth considering when you want a consumer-oriented Windows GUI with commercial support, or when you need video enhancement features beyond mosaic restoration. UniFab also handles upscaling, HDR conversion, and frame interpolation, making it more versatile for general video enhancement outside of mosaic-specific work. If your hardware does not meet Jasna's GPU requirements, UniFab may run on a wider range of configurations.
What it's built on#
- Languages
- C++Python
FAQ#
Is Jasna free to use?
The base application with BasicVSR++ restoration is free and AGPL-3.0 licensed. Enhanced models (UNet 4x, SD 1.5, and LTX) require a supporter contribution of $15 or $25 USD for a license code. That code activates on up to 3 PCs or cloud workspaces and works offline after activation.
What GPU does Jasna require?
NVIDIA GTX 16-series or RTX 20-series or newer, with driver 610 or higher on Windows or 580 or higher on Linux, and compute capability 7.5 or above. AMD RX 7000/9000 (and Ryzen AI) support is experimental with some features unavailable. The LTX model requires an NVIDIA RTX GPU with at least 8 GB VRAM and 16 GB RAM.
How do I install Jasna?
Download the pre-built release from the GitHub releases page for your OS and GPU type. Unpack it into a folder with only ASCII characters in the path. On Windows, double-click jasna.exe; on Linux NVIDIA, run the jasna binary; on Linux AMD, run run_jasna_amd.sh. No Python environment or package manager setup is required for the release build.
How does Jasna compare to UniFab Video Enhancer?
Jasna performs the same core function as UniFab Video Enhancer and is AGPL-3.0 licensed with a free base tier, while UniFab is a commercial product requiring a per-license purchase. Jasna runs entirely on local hardware with no file uploads to external servers. UniFab offers a broader set of video enhancement features (upscaling, HDR) beyond mosaic restoration.
Can Jasna process still images as well as video?
Yes, for supporters. The SD 1.5 model handles still image restoration and is available with a $15 USD or more supporter contribution. The free BasicVSR++ model processes only video files; still image support requires the supporter tier. The CLI also accepts still image inputs when SD 1.5 is activated.
Similar open-source tools#
SOMA
121 GitHub starsParametric 3D modeling via CLI agent: describe, generate, iterate.
Alternative to
Zoo.dev
LTX Desktop
2,096 GitHub starsLocal AI video generation with LTX models on your own hardware
Alternative to
LTX Studio
Wan2.2
17,975 GitHub starsOpen source video AI with text, image, and speech-to-video generation
Alternative to
Kling AI
OpenCreator
13,637 GitHub starsLocal AI workspace for video, image, voice, and translation
Alternative to
HeyGen
Artcraft
30,700 GitHub starsOpen source AI studio for artists who want real control
Alternative to
Runway