v0.5.0

Check also v0.6.0 alpha version with streaming support, stash integration & improved blending for large mosaics:
https://github.com/Kruk2/jasna/releases/tag/v0.6.0-alpha4

🚀 If you want to support this project buy me a coffee 🚀

Remember to copy model_weights folder from alpha11 version to new to avoid engines compilation.
If you have engines from older version, remove them as they have to be recompiled.
Download either windows or linux build.
Alternative windows download: https://gofile.io/d/kdcZ9P

Changelog: v0.4.x → v0.5.0

🚀 New Features

• YOLO detection models — add support for Lada YOLO models (v2 and v4 fast).
• Nvidia RTX Super-Resolution — much faster than TVAI with ok quality. Free! Supports 2x/4x scale selection. (windows only)
• TVAI reimplementation — rewritten from scratch with persistent ffmpeg workers, async push/pop pipeline, and balanced speed vs VRAM usage.
• RF-DETR v5 detection model — v5 recommended threshold: 0.25
• Multi-GPU support
• Working directory option — temp files that jasna produces will be put there
• Benchmark mode (beta)
• Localization — added Thai and Korean languages
• --no-progress CLI flag
• Auto-discover models from model_weights folder

⚡ Performance

• TensorRT sub-engine compilation rework — model compiled into 6 independent TensorRT sub-engines instead of one monolithic unrolled engine:
  • ~70% less engine VRAM (~1.9 GB for clip 60, ~5.4 GB for clip 180)
  • Same engine works for any clip size
  • 2-3x faster than non-compiled PyTorch inference
  • SpyNet fused into sub-engines
• Dynamic VRAM management — usable on low VRAM systems (≤12GB). More VRAM = faster. Adaptive VRAM offloading with decode backpressure.
• PyTorch 2.10 & newer TensorRT

🎨 GUI / UX

• improved tooltips
• lighter font
• improved output path handling
• improved status handling
• improved error messages
• startup screen clamping fix
• job ordering fix
• skip system check if already done