Upscaling models, as ONNX
The model files run by upscale, the command line upscaler, and by the two front ends built on it: Upscale Agent, the desktop app, and UpsizerAI, the web app. All three download from here on demand; nothing needs fetching by hand.
Each model is the original author's weights converted to ONNX by upscale's
python/export_onnx.py: opset and input/output naming are the same across all of them, and
height and width are dynamic so an image can be run whole or in tiles. Most come in two
precisions, <name>.onnx (float32 weights) and <name>_fp16.onnx (float16 weights, float32
input and output), which give the same picture to within one 8-bit level. A few also come as
<name>_int8.onnx: 8-bit weights, turned back to float32 as the network runs, a quarter of the
size and, on most images, within two levels of float32.
Nothing here was trained by this organization. Every file is a conversion, and each remains under the license its author released the original weights under - follow the source link for the terms, which for some models restrict commercial use.
| Model | Scale | fp32 | fp16 | int8 | Architecture | Original weights |
|---|---|---|---|---|---|---|
APISR_GRL_x4 |
x4 | 27.1 MB | 17.5 MB | - | GRL | HikariDawn/APISR |
APISR_RRDB_x2 |
x2 | 18.7 MB | 9.8 MB | - | RRDBNet-6B | HikariDawn/APISR |
AnimeSharpV2_ESRGAN_Soft_x2 |
x2 | 70.0 MB | 36.6 MB | - | ESRGAN | Kim2091/AnimeSharpV2 |
AnimeSharpV2_MoSR_Sharp_x2 |
x2 | 18.0 MB | 9.4 MB | - | MoSR | Kim2091/AnimeSharpV2 |
AnimeSharpV2_MoSR_Soft_x2 |
x2 | 18.0 MB | 9.4 MB | - | MoSR | Kim2091/AnimeSharpV2 |
AnimeSharpV2_RPLKSR_Sharp_x2 |
x2 | 30.8 MB | 16.1 MB | - | RealPLKSR | Kim2091/AnimeSharpV2 |
AnimeSharpV2_RPLKSR_Soft_x2 |
x2 | 30.8 MB | 16.1 MB | - | RealPLKSR | Kim2091/AnimeSharpV2 |
AnimeSharpV3_x2 |
x2 | 70.0 MB | 36.6 MB | - | ESRGAN | Kim2091/AnimeSharpV3 |
BSRGAN_x2 |
x2 | 69.7 MB | 36.4 MB | - | RRDBNet | kadirnar/BSRGANx2 |
MewZoom_V1_x2 |
x2 | 23.2 MB | 12.5 MB | 7.3 MB | MewZoom TrunkNet | andrewdalpino/MewZoom-V1-2X |
MewZoom_V1_UNet_x4 |
x4 | 517.2 MB | 258.9 MB | 129.9 MB | MewZoom U-Net | andrewdalpino/MewZoom-V1-4X-Unet |
RealESRGAN_anime_x4 |
x4 | 18.7 MB | 9.8 MB | - | RRDBNet-6B | xinntao/Real-ESRGAN |
RealESRGAN_x2 |
x2 | 70.0 MB | 36.6 MB | - | RRDBNet | ai-forever/Real-ESRGAN |
RealESRGAN_x4 |
x4 | 69.9 MB | 36.5 MB | - | RRDBNet | ai-forever/Real-ESRGAN |
RealESRGAN_x4plus |
x4 | 69.9 MB | 36.5 MB | - | RRDBNet | xinntao/Real-ESRGAN |
RealESRGAN_x8 |
x8 | 70.0 MB | 36.6 MB | - | RRDBNet | ai-forever/Real-ESRGAN |
RealESR_General_x4 |
x4 | 5.0 MB | 2.5 MB | - | SRVGGNetCompact | xinntao/Real-ESRGAN |
RealWebPhoto_RGT_x4 |
x4 | 92.3 MB | - | - | RGT | Phips/4xRealWebPhoto_RGT |
Swin2SR_Classical_x2 |
x2 | 72.4 MB | 40.0 MB | - | Swin2SR | mv-lab/swin2sr |
Swin2SR_Classical_x4 |
x4 | 73.0 MB | 40.3 MB | - | Swin2SR | mv-lab/swin2sr |
Swin2SR_RealWorld_x4 |
x4 | 72.3 MB | 39.9 MB | - | Swin2SR | mv-lab/swin2sr |
SwinIR_BSRGAN_x4 |
x4 | 55.2 MB | - | - | SwinIR-M | mikestealth/SwinIR |
UltraMix_Smooth_x4 |
x4 | 69.9 MB | 36.5 MB | - | ESRGAN | Kim2091/UltraSharp |
UltraSharp_x4 |
x4 | 69.9 MB | 36.5 MB | - | ESRGAN | Kim2091/UltraSharp |
Using a file directly
https://huggingface.co/baker76/upscale-models/resolve/main/RealESRGAN_x4.onnx
Input is [1, 3, H, W] float32, RGB, 0 to 1; output is the same layout at the model's
scale. The x2 RRDBNet models need even sides, and the window-attention models and MewZoom's
U-Net need sides that are a multiple of 8 or 32; models.json in the UpsizerAI repository records the
multiple, and the peak activation, of each file.