image-adaptive-3dlut

Image-Adaptive 3D LUT — Expert C

This repository contains the best checkpoint from a modern PyTorch adaptation of Learning Image-adaptive 3D Lookup Tables for High Performance Photo Enhancement in Real-time.

The model predicts content-dependent weights for three learnable 33 × 33 × 33 basis LUTs and applies the fused LUT through differentiable trilinear interpolation. It was trained for paired sRGB enhancement using MIT-Adobe FiveK inputs and Expert C targets.

Files

• best.pt: best checkpoint selected by public validation PSNR.
• run_config.json: training configuration and environment summary.
• result.json: final epoch and best validation result.

The checkpoint contains the model state plus training metadata. It is intended for the implementation in WeiChen80percent/AI_photo_editor.

Download and inference

hf download WeiChen80percent/image-adaptive-3dlut best.pt `
  --local-dir image_adaptive_3dlut\trained_model

python -m image_adaptive_3dlut.infer `
  --checkpoint image_adaptive_3dlut\trained_model\best.pt `
  --input path\to\input.jpg `
  --output image_adaptive_3dlut\runs\inference\input_enhanced.jpg

Training result

• Completed epochs: 400
• Train pairs: 3,935
• Public validation pairs: 488
• Best public validation PSNR: 22.3318 dB
• Parameters: 593,516
• Training framework: PyTorch 2.6.0 with CUDA 12.4

Metrics depend on the data split, alignment, resolution and preprocessing and should not be compared directly with results using different settings.

Intended use and limitations

The checkpoint is intended for research, education and photo-enhancement experiments. It learns the visual preference of Expert C and is not a general-purpose restoration model. Results may vary by scene, camera pipeline, color space and input distribution. Inputs are expected to be RGB images scaled to [0, 1].

The MIT-Adobe FiveK dataset is not included. Users are responsible for obtaining it separately and complying with its terms. See the GitHub repository for implementation details, evaluation commands, attribution and third-party notices.