Jasna: GPU-powered mosaic restoration for videos and images
A GPU tool for restoring video mosaics that changes only selected areas and supports live preview.
GitHub Kruk2/jasna Updated 2026-10-19 Branch main Stars 2.4K Forks 187
Python/CUDA/C++ Video restoration BasicVSR++/LTX NVIDIA or AMD GPU

🧭 Decision Guide

Why trending now: Cannot be determined from the provided material.

Try it if you

  • You have an NVIDIA GTX 16-series or RTX 20-series-or-newer GPU and need to restore mosaics in videos.
    README sections “What Jasna does” and “What you need”: BasicVSR++ restores mosaics and requires an NVIDIA GTX 16-series / RTX 20-series or newer.
  • You need to change only mosaic regions rather than reconstructing every frame.
    README sections “What Jasna does” and “How Jasna works”: Segment Editor selects areas, and the result is blended back only into the mosaic area.
  • You need to process VR180 video or view the restoration without creating an output file.
    README section “What Jasna does”: it supports VR180 video and provides live viewing through Video Player.
  • You have an RTX 5090-class GPU and care about BasicVSR++ processing speed.
    README section “Benchmarks”: on an RTX 5090, 1080p H.264 takes 00:34, 3.6x faster than Lada v0.9.1.

Skip it if you

  • You only have a GTX 10-series or older NVIDIA GPU.
    README section “What you need” explicitly says GTX 10-series and older don't work.
  • Your media or installation folder is on a network disk, or its path contains Chinese, Japanese, or other non-English characters.
    README section “What you need”: network disks are unsupported and the path cannot contain non-English characters.
  • You need LTX but do not have an NVIDIA RTX GPU, 8 GB of VRAM, and 16 GB of RAM.
    README section “LTX needs”: LTX requires an NVIDIA RTX GPU, at least 8 GB VRAM, and 16 GB RAM.
  • You accept only free models but require LTX or SD 1.5 functionality.
    README sections “What Jasna does” and “Supporting the project”: LTX and SD 1.5 are supporter models, while BasicVSR++ is free.

Requirements

  • NVIDIA: GTX 16-series / RTX 20-series or newer; driver 610+ on Windows and 580+ on Linux.
  • AMD: Radeon RX 7000/9000 supported by ROCm; Windows requires AMD Adrenalin 26.8.1 or newer.
  • The local disk path cannot contain Chinese, Japanese, accented, or other non-English characters.
  • Windows requires NTFS; Linux supports ext4, Btrfs, XFS, or ZFS; network disks are unsupported.
  • LTX requires an NVIDIA RTX GPU, at least 8 GB VRAM, and 16 GB RAM.

First step (verbatim from README)

run_jasna_amd.sh

Watch out

  • The first NVIDIA job prepares models for the exact GPU and may take 15–60 minutes.
    README section “Quick start”: the first job is slow and model preparation takes about 15–60 minutes.
  • When VRAM is insufficient, Max Clip Size must be lowered, for example from 90 to 60.
    README section “Quick start”: when out of GPU memory, lower Max Clip Size from 90 to 60.
  • After upgrading Jasna, copy the model_weights directory or model preparation may be repeated.
    README sections “Quick start” and “How Jasna works”: prepared models are saved in model_weights and can be copied to a newer version.
  • License codes are available only through Buy Me a Coffee, Unifans, or the Xianyu channel listed in the Chinese README.
    README section “Where to support” and the opening license warning.

Alternatives

  • Lada：Use Lada when lower GPU memory usage matters more; the README says Lada uses less GPU memory, such as 2.0/3.3 GB for 1080p H.264 versus 4.8/5.6 GB for Jasna v0.9.1.
    Benchmarks

Not stated in the README

  • The README does not provide restoration success rates for BasicVSR++ or LTX across different mosaic types.
  • The README does not provide a complete list of supported video codecs, containers, or input/output formats.
  • The README does not provide complete command-line parameters or CLI examples for Windows and Linux.
  • The README does not provide speed or VRAM benchmarks for AMD RX 7000/9000 versus NVIDIA GPUs.
  • The README does not specify the hardware requirements for Video Player live playback at different resolutions and frame rates.
  • The README does not state whether video processing uploads data to a network or runs entirely locally.

💡 Deep Analysis

6
No I use a supported GTX 16-series GPU and a Windows NTFS SSD, but the project must be stored under a path containing Chinese directory names. Will Jasna run correctly?
For: A Chinese-speaking Windows user running Jasna on a local NTFS SSD with a GTX 16-series-or-newer NVIDIA GPU, whose project directory must contain Chinese characters.

No, because the README imposes a hard path-character restriction that cannot be bypassed merely by using compatible hardware and storage.

  • It requires the project folder path to contain no Chinese, Japanese, accented, or other non-English characters.
  • Windows also requires NTFS, and a local SSD is recommended; those conditions address storage and speed, not the path restriction.
  • Quick start instructs users to unpack the release into a folder satisfying the path rule. Initial NVIDIA model preparation may also take 15–60 minutes.

If you cannot move the project and working directories to an English-only path, Jasna does not fit this deployment constraint. The README provides no path-encoding compatibility mode.

  • What you need: “A folder on a local disk, with no Chinese, Japanese, accented or other non-English characters in its path.”
  • More details: “Disk format: NTFS on Windows”
  • Quick start: “Unpack it into a folder with no Chinese, Japanese, accented or other non-English characters in the path.”
Not stated in the README：The README does not clarify whether non-English characters are forbidden only in project and work directories or also in input video paths.；It does not state whether symbolic links or environment variables can work around the restriction.
No I use a Radeon RX 7000 on Windows with AMD Adrenalin 26.8.1, but I also need LTX and secondary restoration models. Can Jasna meet this requirement?
For: A user with a Radeon RX 7000/9000 on Windows using AMD Adrenalin 26.8.1 or newer who wants to avoid buying an NVIDIA GPU.

No, because AMD support is experimental, while LTX and secondary restoration are explicitly NVIDIA-only.

  • The README lists Radeon RX 7000/9000 as experimental and requires AMD Adrenalin 26.8.1 or newer on Windows.
  • It explicitly states: “Secondary restoration and LTX are NVIDIA only,” so expansion paths such as UNet 4x and LTX cannot meet your hardware requirement.
  • AMD is started through run_jasna_amd.sh, and the GPU must be supported by ROCm.

If you only need the basic BasicVSR++ path, an RX 7000 may still be viable. But because your requirement includes LTX and secondary restoration, Jasna does not fully fit.

  • What you need: “Or AMD Radeon RX 7000/9000 (experimental).”
  • What you need: “On Windows, AMD Adrenalin driver 26.8.1 or newer.”
  • More details: “Secondary restoration and LTX are NVIDIA only.”
run_jasna_amd.sh
Not stated in the README：The README does not provide per-model speed, VRAM usage, or encoding compatibility for RX 7000 cards.；It does not state whether AMD supports the same complete detection, preview, and export feature set as NVIDIA.
It depends I have an RTX GPU with at least 8 GB VRAM and 16 GB RAM, and I am willing to pay for supporter models, but I cannot accept LTX being much slower than BasicVSR++. Should I choose it?
For: A creator with an RTX 20-series-or-newer NVIDIA GPU, at least 8 GB VRAM, and 16 GB RAM who wants richer local detail with LTX.

It depends: LTX suits users who prioritize detail over speed, but not strict real-time or high-throughput workflows.

  • The README describes LTX as “slow but more detailed” and explicitly says it is much slower than BasicVSR++.
  • LTX requires an NVIDIA RTX GPU, at least 8 GB VRAM, and 16 GB RAM. Supporter access also requires a license code; the $25 tier includes LTX and future updates.
  • Trial mode performs a speed test showing whether the GPU can run LTX and how long it may take, but it does not restore the video.

If fast batch completion is mandatory, BasicVSR++ is the better fit. If detail in individual shots matters more, run Trial mode before purchasing.

  • What Jasna does: “slow but more detailed LTX”
  • What you need: “LTX needs an NVIDIA RTX GPU, at least 8 GB VRAM and 16 GB RAM.”
  • Supporting the project: “Run LTX in Trial mode (speed test) on your PC.”
Not stated in the README：The README provides no LTX timing table across resolutions, video lengths, or RTX models.；It does not quantify LTX's real quality improvement or rate of incorrect textures versus BasicVSR++.
Yes I process videos through the command line on Linux NVIDIA using an ext4 SSD. How can I avoid waiting another 15–60 minutes for model preparation after upgrading Jasna?
For: A Linux NVIDIA user maintaining a command-line automation workflow on a local ext4 SSD who wants to avoid re-preparing GPU models after Jasna upgrades.

Yes, because the README explicitly documents model-cache reuse for upgrades, and Linux NVIDIA, ext4, and local SSD are supported configurations.

  • Quick start says NVIDIA prepares models for the exact GPU on the first run, taking about 15–60 minutes and only once.
  • When updating Jasna, copying the model_weights folder to the new version avoids waiting again.
  • Linux NVIDIA is supported, ext4 is an accepted filesystem, SSD is recommended, and the project provides a command line suitable for automation.

The cache is tied to the “exact GPU,” however, so the README does not guarantee that old weights remain reusable after changing GPUs or across all version changes.

  • Quick start: “This takes about 15–60 minutes, only once.”
  • Quick start: “When you update Jasna, copy the `model_weights` folder to the new version”
  • More details: “ext4, Btrfs, XFS or ZFS on Linux”
  • Jasna introduction: “has a simple app and a command line”
Not stated in the README：The omitted Command line section does not provide the exact batch command, parameter format, or exit codes.；The README does not say whether model_weights remains compatible with every later version or when redownloading is required.
Yes I use an RTX 5090 and need to batch-restore dynamic mosaics in 720p, 1080p, and 2160p videos while leaving unselected areas unchanged. Is Jasna suitable?
For: A post-production editor using an RTX 5090 who processes 720p, 1080p, and 2160p videos and needs to restore only mosaicked areas without changing the background.

Yes, because Jasna is designed for local mosaic restoration rather than full-frame regeneration.

  • The README states: “Finds mosaics in a video and restores them. The rest of the picture is not touched,” and the Segment Editor allows selecting specific restoration areas.
  • BasicVSR++ is the free, faster model. On an RTX 5090, the benchmark reports about 32, 34, and 63 seconds for a 3:22 video at 720p, 1080p, and 2160p.
  • Video detection, temporal restoration, and the Video Player preview are suitable for checking region boundaries before batch processing.

However, the benchmark covers one GPU, version, and test material, so the same speed cannot be assumed for every codec or moving subject.

  • What Jasna does: “The rest of the picture is not touched.”
  • What Jasna does: “Restores only the parts you choose, with a preview, in the Segment Editor.”
  • Advanced > Benchmarks: RTX 5090 results of 00:32, 00:34, and 01:03 for 720p, 1080p, and 2160p
Not stated in the README：The README does not specify batch concurrency, queue management, or long-run stability.；It does not quantify detector accuracy for fast motion, heavy compression, or complex occlusion.
It depends I need to process 8K VR HEVC 60 fps VR180 footage and want local mosaic restoration on an NVIDIA GPU. Is Jasna worth adopting?
For: An immersive-video creator processing VR180 or 8K VR HEVC 60 fps footage on a local SSD with a recent NVIDIA GPU.

It depends: Jasna matches the VR180 use case, but 8K memory and encoding risks require separate validation.

  • The README explicitly supports VR180 videos, so the project is not limited to standard flat video.
  • Its benchmark includes “8K VR HEVC 60 fps (15 s)”; an older version records about 17.1/18.3 GB of GPU memory usage, showing that this input can be extremely demanding.
  • The project requires a local disk, recommends an SSD, and does not support network disks. NVIDIA also requires compute capability 7.5 or higher.

A high-VRAM NVIDIA GPU and local SSD make adoption plausible. With limited VRAM or a requirement to process long clips reliably, the README does not provide enough evidence.

  • What Jasna does: “Works with VR180 videos”
  • Advanced > Benchmarks: “8K VR HEVC 60 fps (15 s)”
  • GPU memory table: 17.1/18.3 GB recorded for 8K VR HEVC 60 fps in an older version
  • What you need: “Network disks are not supported. Use an SSD.”
Not stated in the README：The README does not give current v0.11.2 processing time or peak VRAM for 8K VR HEVC 60 fps.；It does not state whether VR180 metadata, field-of-view information, and audio/video sync are always preserved.

✨ Highlights

  • Free BasicVSR++ is fast; supporter-only LTX adds detail but requires NVIDIA
  • Segment Editor restores selected areas while leaving the rest untouched
  • On RTX 5090, 1080p restoration takes about 34 seconds, 3.6x faster than Lada
  • The first NVIDIA model preparation takes 15–60 minutes and runs only once
  • LTX requires an NVIDIA RTX GPU, at least 8 GB VRAM, and 16 GB RAM

🔧 Engineering

  • Jasna detects mosaics with rfdetr-v6, then restores clips with BasicVSR++
  • Video Player previews restored video live without first creating a file
  • It supports VR180 video and still images; SD 1.5 requires supporter access

⚠️ Risks

  • GTX 10-series and older GPUs are unsupported; AMD RX 7000/9000 remains experimental
  • The project path cannot contain Chinese, Japanese, accented, or other non-English characters
  • Network disks are unsupported; Windows needs NTFS and Linux needs ext4 or another listed local filesystem
  • Supporter licenses come only from listed channels; Jasna 0.10 keys no longer work

👥 For who?

  • Developers with RTX 20-series or newer GPUs who need video mosaic restoration
  • Users of NVIDIA video tools who need VR180 processing or live preview
  • Individuals willing to use free BasicVSR++ or purchase supporter access to LTX