v4.0.1

[HolyWu]

• Switch to PyTorch again for inference. A few parameters are added and some parameters are removed.
• Add official ESRGAN x4 model and realesr-general-x4v3 model.

See Discussions for benchmarks of some models.

---

This discussion was created from the release v4.0.1.

[HolyWu]

Benchmark

Configuration: NVIDIA RTX 3050, driver 527.37, Windows 10 21H2, Python 3.10.8, VS R61, 720x480 FP16
Data format: FPS / VRAM usage

RealESRGAN_x2plus

Args	Result
num_streams=1	2.69 fps / 832,244 K
num_streams=2	2.85 fps / 1,446,644 K
num_streams=1, nvfuser=True, cuda_graphs=True	2.82 fps / 860,908 K
num_streams=2, nvfuser=True, cuda_graphs=True	2.89 fps / 1,555,192 K
num_streams=1, trt=True	4.52 fps / 596,708 K
num_streams=2, trt=True	4.74 fps / 1,090,284 K

RealESRGAN_x4plus

Args	Result
num_streams=1	0.68 fps / 2,587,388 K
num_streams=2	0.70 fps / 4,956,944 K
num_streams=1, nvfuser=True, cuda_graphs=True	0.71 fps / 2,497,272 K
num_streams=2, nvfuser=True, cuda_graphs=True	0.73 fps / 4,827,924 K
num_streams=1, trt=True	1.18 fps / 1,942,252 K
num_streams=2, trt=True	1.21 fps / 3,783,424 K

RealESRGAN_x4plus_anime_6B

Args	Result
num_streams=1	2.00 fps / 2,562,812 K
num_streams=2	2.12 fps / 4,932,368 K
num_streams=1, nvfuser=True, cuda_graphs=True	2.07 fps / 2,409,204 K
num_streams=2, nvfuser=True, cuda_graphs=True	2.24 fps / 4,674,312 K
num_streams=1, trt=True	3.58 fps / 1,876,716 K
num_streams=2, trt=True	3.81 fps / 3,652,344 K

realesr-animevideov3

Args	Result
num_streams=1	9.68 fps / 387,820 K
num_streams=2	13.01 fps / 590,576 K
num_streams=1, nvfuser=True, cuda_graphs=True	14.41 fps / 287,460 K
num_streams=2, nvfuser=True, cuda_graphs=True	23.45 fps / 439,012 K
num_streams=1, trt=True	14.20 fps / 361,184 K
num_streams=2, trt=True	19.97 fps / 621,284 K

[dan64]

Using RealESRGAN_x4plus_anime_6B (my preferred model) I found that the optimal value is num_streams=3, probably this value represent the safest value that can be used with 12GB.

[dan64]

Hello HolyWu, I'd like to know if you have a plan to add x2 model for RealESRGAN_x4plus_anime_6B. It could be useful to speed up the encoding when a 2x multiplier is enough (for example a DVD source).

[HolyWu]

I can't, because the author of Real-ESRGAN only published x4 model of RealESRGAN_x4plus_anime_6B.

[dan64]

Hello HolyWu, the author changed the model to support also 2x, see the answer at the following link: https://github.com/xinntao/Real-ESRGAN/issues/523#issuecomment-1367535597

[HolyWu]

The guy replied to you (TechnoMasterBoy) is not the author. The author is xinntao. I bet what TechnoMasterBoy did is simply upscaling to 4x first and then downscaling to 2x afterwards.