jasna

Product Introduction

1. Definition: Jasna is a specialized, GPU-accelerated video and image restoration tool designed for the automated detection and repair of mosaic censorship (pixelation) in visual media. It operates as a local desktop application and command-line interface (CLI).
2. Core Value Proposition: Jasna exists to automate the complex, manual process of video inpainting and restoration, specifically targeting mosaic degradation. Its primary value is enabling archivists, content creators, and media professionals to efficiently restore historical footage, home movies, or censored content with high-quality, AI-driven results directly on their local hardware, ensuring privacy and control.

Main Features

1. AI-Powered Mosaic Detection & Tracking: Utilizes a custom-trained object detection model (rfdetr-v6) to automatically identify mosaic regions in each video frame. It then employs temporal tracking to group these regions across frames into coherent "clips," ensuring consistent restoration across shots and stopping at scene cuts to prevent cross-contamination.
2. Multi-Model Restoration Pipeline: Features a two-stage, GPU-optimized restoration core. The primary, free stage uses BasicVSR++, a video super-resolution model, to restore detected 256x256 mosaic clips. An optional secondary stage can apply supporter-only models like UNet 4x for sharpening or the experimental SD 1.5 (Stable Diffusion 1.5) for image restoration, enhancing detail.
3. Hardware-Accelerated End-to-End Processing: Leverages GPU hardware encoders/decoders (NVIDIA NVDEC/NVENC or AMD AMF) for full pipeline acceleration, from video decoding and encoding to AI inference. This design minimizes CPU overhead and maximizes throughput for fast video restoration.
4. Precision Editing with Segment Editor: Includes an interactive editor that allows users to review auto-detected mosaics, manually add or remove areas for restoration, and apply different models (e.g., LTX only on large mosaics) to specific segments, providing granular control over the output.
5. Supporter-Exclusive LTX Diffusion Model: A premium, NVIDIA-only feature for supporters. LTX uses a multi-pass, diffusion-based approach that is significantly slower than BasicVSR++ but is capable of generating highly detailed and coherent textures, offering a higher-quality restoration tier for demanding use cases.

Problems Solved

1. Pain Point: The labor-intensive and skill-dependent nature of manually restoring pixelated or mosaicked areas in video and images using traditional editing software like Adobe After Effects or DaVinci Resolve.
2. Target Audience: Video archivists and preservationists, independent filmmakers working with degraded source material, media analysts, and enthusiasts seeking to restore personal or historical censored video content.
3. Use Cases: Restoring old VHS tapes or DVDs with compression artifacts masked by mosaics; processing censored footage for archival or analytical purposes; enhancing specific degraded regions in VR180 content; batch processing folders of images affected by pixelation.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros: Exceptional processing speed due to full GPU pipeline optimization; focused, non-destructive restoration that only alters mosaic areas; offers a powerful free tier (BasicVSR++); includes advanced interactive tools like the Segment Editor and live Video Player preview.
   • Cons: High GPU requirements (NVIDIA GTX 16-series/RTX 20-series or newer with compute capability 7.5+); limited AMD support (experimental, RX 7000/9000 series only); premium models (LTX, UNet 4x) require a paid license; pathnames cannot contain non-ASCII characters, which is restrictive for some international users.

2. Key Alternatives & Differentiation:

   • vs. Topaz Video AI: Topaz is a broader video enhancement suite (upscaling, deinterlacing, motion deblur). Jasna is hyper-specialized in mosaic inpainting and detection, making it more effective and faster for this specific task. Jasna's free tier also provides core functionality without Topaz's high one-time cost.
   • vs. Manual Inpainting (e.g., with EbSynth, Stable Diffusion): Manual workflows offer ultimate control but are extremely time-consuming and require significant expertise. Jasna automates detection, tracking, and batch processing, reducing a multi-hour manual task to a mostly automated process with comparable or superior quality for its specific purpose.
   • vs. Lada (Inspiration): Jasna is directly inspired by Lada and uses a model trained by its author. The key differentiation is performance: Jasna's fully GPU-accelerated pipeline makes it 3-4x faster than Lada in benchmarks, while also offering a more polished GUI and supporter models like LTX.

Frequently Asked Questions (FAQ)

1. Is Jasna completely free to use? Yes, the core application with the BasicVSR++ restoration model is completely free. Advanced models like LTX (for higher-detail video restoration) and UNet 4x/SD 1.5 (for sharpening and image restoration) require a one-time supporter license purchased through official channels like Buy Me a Coffee.
2. What are the minimum GPU requirements for Jasna? You need an NVIDIA GPU from the GTX 16-series or RTX 20-series onward (e.g., RTX 2060, 3060, 4060) with a recent driver. Older GPUs (GTX 10-series) are not supported. Experimental AMD support exists for Radeon RX 7000/9000 series and recent Ryzen AI iGPUs.
3. Can Jasna restore videos that are not mosaicked? No, Jasna is specifically engineered for detecting and inpainting mosaic/pixelated regions. It is not a general-purpose video enhancer for noise, blur, or low resolution unless those issues are part of a mosaic area it is restoring.
4. Why is the first run of Jasna so slow? On first launch with an NVIDIA GPU, Jasna compiles and optimizes its AI models (TensorRT engines) specifically for your exact GPU architecture. This one-time process can take 15-60 minutes but results in significantly faster performance for all subsequent jobs.
5. How does Jasna handle audio and video formats? It uses your GPU's hardware decoder and encoder, supporting common codecs like H.264 and HEVC. Audio is typically copied directly from the source stream without re-encoding, preserving original quality, unless the output container forces a re-encode.