Inside Jasna’s low-VRAM video restoration pipeline

Jasna, a Python project for reconstructing mosaic-censored regions in Japanese adult videos and images, has passed 2,300 GitHub stars. Its narrow application combines several broadly useful inference techniques: a transformer detector ahead of an inpainting network, fixed-batch TensorRT engines, automatic GPU memory management, and scene-aware temporal processing.

The code builds on Lada and uses mosaic_restoration_1.2, a model trained by Lada’s author. Jasna contributes the surrounding engineering, including a faster detection stage, lower VRAM consumption, side-by-side VR180 support, interactive segment editing, and optional secondary upscalers.

Detection first, restoration second

Jasna detects censored regions in each frame, then sends those regions through the restoration model in temporal clips. RF-DETR v6 is the default detector, while larger RF-DETR and YOLO variants trade additional compute for detection quality.

Video-specific controls prevent adjacent frames from being processed as unrelated images. Temporal overlap and crossfades reduce flicker at clip boundaries, while hard-cut detection stops the system from blending restoration data across scene changes.

• Segment editor: Previews footage, limits processing to selected ranges, and supports zooming and panning.
• Native player: Seeks through restored frames without first writing a completed output file.
• Streaming mode: Sends processed video to a browser player or a compatible Stash fork.
• VR180 handling: Processes side-by-side footage per eye and supports studio-specific restoration settings.

Fixed engines shrink the memory curve

The repository reports its main benchmarks on an RTX 5090 paired with an Intel Core i9-13900K. The figures compare Jasna with Lada under project-defined test conditions, so codec settings, model selection, driver versions, and input characteristics remain relevant when reproducing them.