- Fix ESRGAN, FDAT, HAT and DAT models using more GPU memory on the native GPU kernels than on TensorRT: they now use about half as much as before, less than TensorRT
| 1080p video |
Speedup: NVIDIA RTX 5090 |
Speedup: RTX 4090 |
Speedup: RTX A5000 (RTX 30) |
VRAM: TensorRT building its engine |
VRAM: TensorRT |
VRAM: native |
| 2x Performance |
1.7x |
1.7x |
1.8x |
1.85 GB |
1.06 GB |
0.93 GB (13% less) |
| 2x Balanced |
1.6x |
1.4x |
1.15x |
1.94 GB |
1.20 GB |
1.02 GB (15% less) |
| RIFE interpolation |
1.9x |
3x |
2.5x |
1.83 GB |
1.12 GB |
0.82 GB (27% less) |
| TemporalFix |
1.4x |
2.1x |
2.2x |
2.13 GB |
1.85 GB |
1.42 GB (24% less) |
| MangaJaNai 2x |
1.6x |
Not measured |
Not measured |
4.54 GB |
3.73 GB |
2.67 GB (28% less) |
| IllustrationJaNai FDAT-M 4x |
1.7x |
Not measured |
Not measured |
9.14 GB |
7.33 GB |
6.75 GB (8% less) |
| IllustrationJaNai HAT-L 4x |
1.6x |
Not measured |
Not measured |
15.58 GB |
15.59 GB |
10.41 GB (33% less) |
| IllustrationJaNai DAT-2 4x |
2.3x |
Not measured |
Not measured |
12.25 GB |
12.25 GB |
10.90 GB (11% less) |
VRAM was measured on the RTX 5090. Native's savings are against TensorRT once its engine is built; native GPU kernels have no engine to build. TensorRT can't build HAT-L for 1080p, so its HAT-L figures are for the frame padded to 1920x1088, as the native kernels pad it.
Downloads