Texel-4.5 (fast 2Γ— upscaler for textures and photos)

Texel-4.5 is the fast Texel model. It's a 2Γ— upscaler that removes compression damage (BC1/DXT block grids, JPEG blocking, noise) and rebuilds fine surface detail, and it runs about 1.35–1.5Γ— faster than Real-ESRGAN x2plus with half the VRAM.

It beats Real-ESRGAN x2plus in every quality category, on every test set and on every input type we tested (see below). If you want the most detail and speed doesn't matter, Texel-4.2 still scores higher on quality overall. Texel-4.5 is about twice as fast as 4.2.

Download one file: Texel-4.5_2x_fp32.onnx (88 MB). It works with any input size (odd sizes too), and the output is always exactly 2Γ—.

Using it

chaiNNer (no Python)

  1. Download Texel-4.5_2x_fp32.onnx.
  2. In chaiNNer, use Load Model (ONNX) β†’ Upscale Image.

onnxruntime (Python)

import numpy as np, onnxruntime as ort
from PIL import Image

sess = ort.InferenceSession("Texel-4.5_2x_fp32.onnx",
                            providers=["CUDAExecutionProvider", "CPUExecutionProvider"])
img = np.asarray(Image.open("in.png").convert("RGB"), np.float32) / 255.0
x = img.transpose(2, 0, 1)[None]                 # 1x3xHxW, RGB, 0..1
y = sess.run(None, {"input": x})[0][0]           # 3x2Hx2W
out = (y.transpose(1, 2, 0).clip(0, 1) * 255 + 0.5).astype(np.uint8)
Image.fromarray(out).save("out_2x.png")
  • Input input: float32, [batch, 3, H, W], RGB in 0..1. Output output: [batch, 3, 2H, 2W].
  • RGB only. For textures with transparency, upscale the alpha channel separately (for example with Lanczos).
  • Tiling textures: the ONNX file is not tile-aware. For seamless repeating textures, wrap-pad the input by 16 px on every side, upscale it, then crop 32 px off each edge of the output.
  • Use it on colour (albedo) textures and photos. Don't run it on normal maps, masks or other data maps.

Speed

RTX 4070 Ti SUPER, 2Γ— upscale, whole image in one pass, median GPU time after warm-up (model only, no file loading or saving). PyTorch fp16 + torch.compile.

Input Texel-4.5 Real-ESRGAN x2plus Texel-4.2
256Γ—256 11 ms 40 ms 22 ms
512Γ—512 40 ms 62 ms 93 ms
1024Γ—1024 (texture) 170 ms 260 ms 365 ms
1920Γ—1080 (1080p frame) 318 ms 429 ms 723 ms
VRAM at 1080p 1.19 GB 2.46 GB 2.49 GB

In plain fp32 the gap is the same: 818 ms for Texel-4.5 vs 1091 ms for Real-ESRGAN on a 1080p frame.

Quality

Measured with the Texel Visual Benchmark (TVB) on 444 test cases: held-out Breath of the Wild textures, held-out Poly Haven textures and DIV2K photos, each fed in clean, as JPEG q75 and as BC1. No contestant was trained on any test image.

Each image is shrunk 2Γ—, upscaled back and compared with the original over 9 visual metrics, grouped into detail, artifacts, structure and colour. The Visual Score is the percentage of head-to-head comparisons won against all the other models (50 = average). An image with a clearly visible artifact loses outright to one without.

Overall

# Model Visual Score Detail Artifacts Structure Colour Visible flaw PSNR
1 Texel-4.2 71.7 84.0 58.5 66.0 68.2 2% 29.89 dB
2 Texel-4.5 65.0 74.0 61.7 54.0 61.1 5% 29.26 dB
3 Texel-4.1 61.8 76.2 46.9 56.4 55.3 13% 29.68 dB
4 Texel-4 55.4 38.5 58.9 77.4 65.3 11% 31.02 dB
5 Lanczos 45.7 22.3 68.9 51.4 61.9 6% 29.61 dB
6 SwinIR-M 41.4 42.1 37.0 44.9 42.5 41% 30.45 dB
7 Real-ESRGAN x2plus 40.3 44.4 57.3 22.0 25.6 5% 27.39 dB
8 HAT 36.4 36.5 32.0 41.2 37.2 50% 30.53 dB
9 HAT-L 32.3 32.0 28.7 36.8 33.0 58% 30.52 dB

Texel-4.5 vs Real-ESRGAN x2plus, per test set and input

Test Texel-4.5 Real-ESRGAN x2plus Texel-4.2
BotW textures (held out) 63.3 36.5 73.5
Poly Haven textures (held out) 76.8 (1st of 9) 46.5 76.4
DIV2K photos 61.1 38.7 69.8
Clean input 56.9 22.5 64.4
JPEG q75 input 69.4 53.1 77.8
BC1 input 68.7 45.4 72.8

Texel-4.5 also leaves the fewest compression blocks of any model tested (raw blocks score 0.024, vs 0.035 for Real-ESRGAN and 0.197 for Texel-4.2).

Model

  • Architecture: UnshuffleSR, 22.0M parameters. The input is pixel-unshuffled so the main body (192 channels, 64 convolutions) runs at half resolution, then a full-resolution refine stage and PixelShuffle produce the 2Γ— output, plus a nearest-neighbour skip. Working at half resolution is what makes it fast.
  • Training, all on Breath of the Wild and Bonelab colour textures plus CC0 Poly Haven material scans, with a random mix of downscaling, JPEG, BC1, noise and banding on the inputs:
    1. Base: 60k steps from scratch, pixel L1 plus distillation from Texel-4.2-Base.
    2. Detail: four fine-tunes in a row (100k steps in all), distilled from Texel-4.2 with weight 2–4. A per-phase edge-profile penalty stops the upsampler from leaving a 2/4/8/16 px repeating rhythm or block grid, and an FFT penalty stops invented pixel-grid patterns.

Limitations

  • Overall quality is below Texel-4.2 (65.0 vs 71.7), mostly on structure (edges exactly where they belong). Use 4.2 when you want the best result and don't mind it being about twice as slow.
  • It can add grain. On textures that already have fine noise, like grass or gravel, it can over-sharpen and add a crunchy grain. If that bothers you, blend its output with a plain Lanczos upscale, or use Texel-4.2.
  • It's tuned for detail, so some of the detail it adds is plausible rather than the exact original. For the closest pixel-for-pixel match, use a PSNR-trained model such as Texel-4.

License

The model weights are released under CC0 1.0 (public domain dedication).

It was trained on textures from The Legend of Zelda: Breath of the Wild and Bonelab, CC0 textures from Poly Haven, and DIV2K photos. None of that training data is included here, and its copyright stays with its owners. This project isn't affiliated with or endorsed by Nintendo, Stress Level Zero or Poly Haven.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support