On NVIDIA RTX 30, 40 and 50 GPUs the AnimeJaNai models now run on their own GPU kernels instead of TensorRT: upscaling is up to 1.8x faster, RIFE interpolation up to 3x, and there's no more waiting for a TensorRT engine to build, so a new profile starts playing in seconds. Frames that don't change, like held drawings and still shots, are no longer upscaled twice, which cuts GPU power by up to 40%, so your GPU runs cooler and quieter.
| Speedup over TensorRT, 1080p | NVIDIA RTX 5090 | RTX 4090 | RTX A5000 (RTX 30) |
|---|---|---|---|
| 2x Performance | 1.7x | 1.7x | 1.8x |
| 2x Balanced | 1.6x | 1.4x | 1.15x |
| RIFE interpolation | 1.9x | 3x | 2.5x |
| TemporalFix | 1.4x | 2.1x | 2.2x |
- Upscale on native GPU kernels on NVIDIA RTX 30 and newer GPUs: faster on every model and GPU tested, with no TensorRT engine to build first
- Reuse the upscale of unchanged frames such as held drawings and still shots, for up to 40% less GPU power at 1080p to 4K (AnimeJaNai Manager > Reuse upscales of unchanged frames, on by default)
- Show the share of reused frames and whether native kernels are running in the stats overlay (Ctrl+J)
- Turn the native GPU kernels off to use TensorRT engines instead, if a model looks wrong or fails (AnimeJaNai Manager > Use native GPU kernels)
- Fix AnimeJaNai Manager > Components failing to load ("'T' is an invalid start of a value") when GitHub limits requests, and check for updates without using up that limit
- Update mpv to latest upstream
Downloads
| Installer | Portable | |
|---|---|---|
| Windows | Setup .exe | .7z |
| Linux | AppImage | .tar.zst |