Outlier Upscale Method — High quality models
Real-ESRGAN weights converted for the browser, used by the High quality mode of Outlier Upscale Method (an in-browser 2K/4K image upscaler).
| File | Source checkpoint | Architecture | Size |
|---|---|---|---|
realesrgan-x4plus.v1.part1.bin + realesrgan-x4plus.v1.part2.bin |
RealESRGAN_x4plus | RRDBNet, 23 blocks, ×4 | 33.4 MB |
realesrgan-x4plus-anime6b.v1.bin |
RealESRGAN_x4plus_anime_6B | RRDBNet, 6 blocks, ×4 | 8.9 MB |
The x4plus file is split in two only so it also fits hosts with a 25 MiB per-file limit; the page downloads both parts and joins them byte for byte.
Format
Float16 little-endian, no header. Convolutions in order (conv_first, every body.i.rdbN.convM,
conv_body, conv_up1, conv_up2, conv_hr, conv_last), each stored as a TensorFlow filter
[kh, kw, in, out] followed by its bias.
Produced by tools/convert_models.py in the web tool's GitHub repository, directly from the
official checkpoints at https://github.com/xinntao/Real-ESRGAN/releases. Not modified or retrained.
License and credit
Real-ESRGAN © 2021 Xintao Wang, BSD 3-Clause License. Paper: Xintao Wang, Liangbin Xie, Chao Dong, Ying Shan, Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data, ICCV Workshops 2021.
Converted and maintained by Truong Dang Khoa (Outlier Code).