Kruk2/jasna

The problem

Removing censor mosaics from a video is usually a manual job: mask each frame, paint over it, re-encode. Frame-by-frame segmentation and tracking is tedious, and generic video inpainting tools treat the whole frame as the target, which risks altering parts that should stay untouched. Jasna automates the loop: it decodes on the GPU, detects mosaics per frame, groups each mosaic into clips, restores each clip with a sequence-aware model, and blends only the restored patch back into the original.

What's different

The project is built around scoped restoration rather than whole-frame enhancement. Detection, tracking and clip boundaries come first, so the model only ever sees a 256x256 patch and the untouched pixels keep their original quality. Restoration is split into two tiers: a free BasicVSR++ path and a slower diffusion-based LTX path for supporters, with optional secondary upscaling (UNet 4x, RTX Super Res, Topaz TVAI). Compared with general-purpose video inpainting frameworks, the trade-off is narrowness: it is purpose-built for one task and one GPU stack, not a platform you can retarget. A preview player that restores live without writing a file is a genuinely useful workflow detail that most pipelines lack.

Who it's for

Individual users with an NVIDIA GTX 16-series/RTX 20-series or newer (AMD RX 7000/9000 is labelled experimental) who process local video files and want a GUI plus a CLI. It is not for CPU-only machines, older GPUs, cloud/network storage, or paths containing non-English characters. Teams needing batch orchestration, reproducible pipelines, or permissive licensing should look elsewhere. Anyone uncomfortable with a paid-license gate on the better models should stick to the free tier or skip it.

Cost of adoption

Hardware is the real barrier: NVIDIA compute capability 7.5+, recent drivers, SSD storage on NTFS or a Linux filesystem, and for LTX at least 8 GB VRAM and 16 GB RAM. The first run spends roughly 15-60 minutes compiling models for your GPU, and that model_weights folder must be carried forward on upgrades. The app is free, but the sharper UNet 4x, SD 1.5 image model and LTX are unlocked by a one-time supporter payment ($15-$25), tied to a license code valid on 3 machines with online activation. The license is AGPL-3.0, so redistributing a modified version has source obligations. Installation is by downloading a release rather than a package manager; follow the README.

Verdict

Jasna is a young project, created in early 2026, with about 3.5k stars and a single weekly board appearance of +1,469, so treat current attention as a spike rather than a proven track record. The engineering approach is coherent and the GUI/CLI split is more polished than most single-purpose GPU tools, but the supporter-license model and strict hardware matrix are real commitments. Adopt it if you have the right GPU, a local workflow and this specific need; otherwise the free tier is a cheap way to test the waters. Expect churn and read the licensing terms carefully before building anything around it.