CANDLE

Color Ambient Normalization with DINO Layer Enhancement

Overview

CANDLE is a color ambient normalization framework built on multi-stage DINO feature guidance and a coarse-to-refine restoration pipeline.

This repository contains the training and inference code for:

• backbone training
• refine-stage training
• joint coarse + refine training
• inference and test-time augmentation
• DINO token extraction

Core method elements:

• multi-stage DINO layer fusion
• query-guided feature conditioning
• coarse-to-refine restoration
• ambient light normalization

Environment

Install dependencies with:

pip install -r requirements_aln_final.txt

Recommended environment:

• Python 3.10
• PyTorch with CUDA
• NVIDIA GPU for training and inference

Repository Structure

CANDLE/
├── model.py
├── refine_nafnet.py
├── train.py
├── train_refine.py
├── train_refine_joint.py
├── test.py
├── test_tta.py
├── extract_dino_tokens32.py
├── options.py
├── utils/
└── NAFNet/

Main Entry Points

• train.py: backbone training
• train_refine.py: refine-only training
• train_refine_joint.py: joint coarse + refine training
• test.py: inference
• test_tta.py: inference with test-time augmentation
• extract_dino_tokens32.py: DINO token extraction

Key Files

• model.py: CANDLE backbone
• refine_nafnet.py: refinement network
• utils/aln_dataset.py: dataset loader
• utils/val_utils.py: validation metrics
• utils/schedulers.py: scheduler helpers

Data Preparation

Training and inference expect:

• input images
• ground-truth target images
• pre-extracted DINO token files

The DINO features can be prepared with:

python extract_dino_tokens32.py \
  --input_dir /path/to/images \
  --out_dir /path/to/dino_tokens32 \
  --target_w 1024 \
  --target_h 768 \
  --batch_size 8

Training

Example entry points:

python train.py ...
python train_refine.py ...
python train_refine_joint.py ...

The final joint training pipeline is implemented in train_refine_joint.py.

Inference

Run inference with:

python test.py ...

For test-time augmentation:

python test_tta.py ...

Supported Optimization Options

• The codebase supports pixel-oriented fine-tuning with L1, MSE, and Charbonnier losses.
• The joint pipeline also supports fixed, cosine, warmup_cosine, and plateau learning-rate schedules.

Repository Scope

This repository is prepared as a code-only package for training and inference.

It intentionally does not include:

• datasets
• training checkpoints
• experiment logs
• visualization outputs

These artifacts are ignored through .gitignore and should be managed separately from the source repository.

Notes

• Update dataset paths and checkpoint paths in your training or inference commands before running.
• Some helper shell scripts are included as references and may need local path adjustment.