mashbox-labs/Real-ESRGAN-TPU

Real-ESRGAN x2plus and x4plus in BSHJ

Weights only: the mashbox package, release 0.1.10, has no mashbox run command for this model.

Xintao Wang's Real-ESRGAN image upscaling weights, 2x and 4x, as BSHJ files, the weight format the mashbox TPU engine reads. Float32, converted from the upstream release files (params_ema).

Files

File	Scale	Size (MB)	Tensors	Converted from
RealESRGAN_x2plus.bshj	2x	68.3	702, float32	RealESRGAN_x2plus.pth, Real-ESRGAN release v0.2.1
RealESRGAN_x4plus.bshj	4x	68.3	702, float32	RealESRGAN_x4plus.pth, Real-ESRGAN release v0.1.0

Tensor names and shapes are the upstream PyTorch state dict's (conv_first.weight, body.0.rdb1.conv1.weight, and so on). Each file's header records the upstream file it came from and that file's sha256.

Load

The mashbox package's BSHJ reader opens the files as NumPy arrays over a memory map, without copying:

pip install mashbox

from huggingface_hub import hf_hub_download
from mashbox import bshj

f = bshj.File(hf_hub_download("mashbox-labs/Real-ESRGAN-TPU", "RealESRGAN_x4plus.bshj"))
print(f.meta()["source"])          # upstream file, its sha256, state dict key
w = f.array("conv_first.weight")   # float32, shape (64, 3, 3, 3)

Links

Original code and weights	xinntao/Real-ESRGAN by Xintao Wang
BSHJ reader	src/mashbox/bshj.py
Companies	hello@mashbox.gg

Licence

BSD-3-Clause, the licence of xinntao/Real-ESRGAN, whose GitHub releases publish these weights ("Copyright (c) 2021, Xintao Wang"). LICENSE in this repo is that repo's licence file, unchanged.