razordvrz/LUTwithBGrid-ONNX

markdown

license: apache-2.0 library_name: onnxruntime tags: - image-enhancement - color-correction - 3d-lut - bilateral-grid - onnx

LUTwithBGrid (ONNX Export)

ONNX export of LUTwithBGrid — Image-Adaptive 3D Lookup Tables for Real-time Image Enhancement with Bilateral Grids (ECCV 2024).

Original

• Paper: Image-Adaptive 3D Lookup Tables for Real-time Image Enhancement with Bilateral Grids
• Authors: Wontae Kim, Nam Ik Cho (Seoul National University)
• Original Repo: WontaeaeKim/LUTwithBGrid
• License: Apache 2.0

What is this

This is a cross-platform ONNX export of the original PyTorch model. The original implementation requires custom CUDA kernels (lut_transform, bilateral_slicing) which must be compiled manually — this can be a barrier for users without CUDA Toolkit or Visual Studio Build Tools.

The ONNX version removes that barrier: it runs on CPU via onnxruntime, no CUDA, no compilation, no PyTorch required.

Changes from Original

1. Rewritten CUDA kernels in pure PyTorch (F.grid_sample + gather) — for ONNX compatibility
2. Exported to ONNX (opset 17) — runs on CPU
3. Hybrid mode — takes brightness from the model, keeps original color (prevents color cast on PBR textures)

Files

File	Description
lutwithbgrid_fivek.onnx	ONNX model (FiveK sRGB weights), 1.9 MB
export_script.py	Script used to export the ONNX model
test_inference.py	Minimal example: load ONNX and run inference
requirements.txt	Python dependencies

Usage

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

# Load model
session = ort.InferenceSession("lutwithbgrid_fivek.onnx")

# Preprocess
img = Image.open("input.png").convert("RGB").resize((512, 512))
arr = np.array(img).astype(np.float32) / 255.0
arr = np.transpose(arr, (2, 0, 1))[None, ...]

# Inference
result = session.run(None, {"input": arr})[0]

# Postprocess
result = np.transpose(result[0], (1, 2, 0))
result = np.clip(result * 255, 0, 255).astype(np.uint8)
Image.fromarray(result).save("output.png")
Intended Use
Image color enhancement / retouching

PBR texture albedo correction

Any workflow where a lightweight, CPU-friendly image enhancement model is needed

Limitations
Trained on FiveK (photographs, not PBR textures) — may produce color casts on synthetic textures

Input size is fixed at 512×512 for training; larger images should be tiled or resized

Hybrid mode (brightness only) is recommended for PBR albedo to avoid color shifts

Citation
bibtex
@inproceedings{kim2024LUTwithBGrid,
  title={Image-adaptive 3D Lookup Tables for Real-time Image Enhancement with Bilateral Grids},
  author={Kim, Wontae and Cho, Nam Ik},
  booktitle={European Conference on Computer Vision},
  year={2024}
}
Acknowledgments
Original model by Wontae Kim and Nam Ik Cho (Seoul National University), licensed under Apache 2.0.

## Used in

This ONNX export powers the **LUTwithBGrid** AI mode in [**Albedolizer**](https://github.com/invisiblelevel/Albedolizer) — a free PBR texture checker & optimizer for 3D artists.

- 🌐 Website: [invisiblelevel.github.io/Albedolizer](https://invisiblelevel.github.io/Albedolizer/)
- 💾 Download: [invisiblelevel.itch.io/albedolizer](https://invisiblelevel.itch.io/albedolizer)
- 🐙 Source: [github.com/invisiblelevel/Albedolizer](https://github.com/invisiblelevel/Albedolizer)

If you use this ONNX model, feel free to also check out Albedolizer for a complete PBR texture workflow.