Most video/image restoration tools try to change entire frames or rely on manual inpainting; Jasna takes the opposite approach: it automatically finds mosaic-censored regions, restores only those crops using temporal video models, and blends results back so the original scene outside the mosaic remains intact. That focus makes it practical for batch-processing long videos while minimizing collateral changes.
What Sets It Apart
- Region-first pipeline: detection → tracking → per-clip restoration → blend. Finds mosaic regions per frame, groups them into clips, restores 256×256 crops with BasicVSR++ and stitches them back to avoid changing non-mosaic pixels.
- Two-tier restoration and previewing: BasicVSR++ provides a fast, free primary pass; LTX (NVIDIA-only, slower) offers higher detail and is available to supporters. Optional secondary upscalers (UNet 4x supporter model, RTX Super-Res, or Topaz Video AI) sharpen large crops before blending.
- GPU-native and playback support: almost everything runs on the GPU (decode/detect/restore/encode), includes a GUI with a live Video Player to preview restored output without writing files, plus a CLI for automation and batch jobs.
- Practical constraints surfaced: initial model preparation for a specific GPU can be lengthy (typically 15–60 minutes on NVIDIA), disk paths must avoid non-ASCII characters, and some features (LTX, secondary restoration, SD 1.5 image inpainting) are NVIDIA-only or supporter-locked.
Who It's For and Trade-offs
Great fit if you need to automatically remove mosaic censorship from many video files or occasional images while preserving the original scene outside the censored regions, and you have a modern GPU (NVIDIA RTX recommended or recent AMD experimental support). It’s also a pragmatic choice when you want previews and a GUI plus CLI for batch workflows.
Look elsewhere if you require a pure CPU solution, need guaranteed cross-platform vendor-agnostic secondary upscalers, or expect out-of-the-box perfection on every still image — still-image restoration uses a Stable Diffusion 1.5 inpaint model (supporter-only) and results are variable; large crops can still benefit from secondary upscaling.
Where It Fits
Jasna sits between lightweight detection-only tools and heavy-handed full-frame inpainting suites: compared with Lada it inherits and extends mosaic-restoration models and adds a more integrated GPU pipeline and preview experience; compared with Topaz or paid commercial tools, Jasna emphasizes reproducible, GPU-accelerated open workflows with optional paid supporter models for extra quality.
