LOREAL: Mitigating Low-Resolution Challenges in Prompt Learning with Attribute-Driven Self-Distillation

Overview

LOREAL is a prompt-learning framework for low-resolution vision-language recognition. It augments a CoOp-style prompt learner with attribute-conditioned prompt tokens and trains them through self-distillation between a standard-resolution branch and a low-resolution branch.

The repository contains:

• CoOp baseline training and evaluation.
• CoOp_LOREAL, the LOREAL trainer built on top of CoOp.
• Low-resolution image handling through Dassl transforms.
• Scripts for staged base-to-new experiments.
• Dataset configs for common prompt-learning benchmarks.

Method

LOREAL keeps the CLIP backbone frozen and learns prompt-side adaptation. Given image features, small meta-networks generate attribute prompt tokens such as color, shape, size, structure, and outline. A standard-resolution student and a low-resolution student are then coupled with:

• Cross-entropy supervision on base classes.
• High-level distillation between prediction distributions.
• Low-level distillation between generated attribute contexts.

At inference time, the low-resolution branch uses image-conditioned attribute prompts for classification.

Code Structure

Path	Description
train.py	Main training and evaluation entry point
trainers/coop.py	CoOp baseline trainer
trainers/coop_loreal.py	LOREAL trainer
configs/trainers/CoOp/	CoOp configs
configs/trainers/CoOp_LOREAL/	LOREAL configs
configs/datasets/	Dataset configs
scripts/run_coop_loreal.sh	General staged LOREAL runner
scripts/run_lr_base_new.sh	Dataset-parameterized low-resolution base/new summary runner
Dassl.pytorch/	Local Dassl fork used by the trainers

Installation

Create an environment with PyTorch installed for your CUDA version, then install the project dependencies:

git clone https://github.com/xuc865/loreal.git
cd loreal

pip install -r requirements.txt
cd Dassl.pytorch
pip install -r requirements.txt
python setup.py develop
cd ..

If CLIP assets are not already cached, the first run will download them automatically.

Datasets

Put datasets under a single root directory, for example:

/path/to/datasets/
  oxford_flowers/
  oxford_pets/
  caltech-101/
  ...

Dataset-specific layouts follow the CoOp/Dassl conventions. See docs/DATASETS.md for dataset names and expected files.

Run

General Runner

Use scripts/run_coop_loreal.sh for staged experiments on any supported dataset:

bash scripts/run_coop_loreal.sh \
  --data-root /path/to/datasets \
  --output-root /path/to/runs \
  --datasets oxford_flowers \
  --resolutions 96 \
  --seeds 1 \
  --stage all

Run multiple datasets, resolutions, and seeds by passing comma-separated lists:

bash scripts/run_coop_loreal.sh \
  --data-root /path/to/datasets \
  --output-root /path/to/runs \
  --datasets oxford_flowers,oxford_pets,caltech101 \
  --resolutions 96,144,192 \
  --seeds 1,2,3 \
  --shots 16 \
  --stage all

Common options:

Option	Meaning
--datasets	Comma-separated dataset keys
--resolutions	Comma-separated low-resolution sizes
--seeds	Comma-separated random seeds
--shots	Number of base-class shots
--stage	all, loreal, stage1, stage2, stage3, or stage4
--dry-run	Print commands without executing them

The staged pipeline is:

1. Train a standard-resolution CoOp teacher.
2. Train a low-resolution CoOp baseline.
3. Train LOREAL self-distillation.
4. Evaluate the low-resolution base/new splits.

Base/New Summary Script

For a compact base/new run with an explicit summary file, use scripts/run_lr_base_new.sh. The script is dataset-parameterized through DATASET:

DATA_ROOT=/path/to/datasets \
OUTPUT_ROOT=/path/to/runs \
DATASET=oxford_flowers \
SEED=1 \
RES=96 \
RESET_LOREAL=1 \
bash scripts/run_lr_base_new.sh

Use another supported dataset by changing DATASET:

DATA_ROOT=/path/to/datasets \
OUTPUT_ROOT=/path/to/runs \
DATASET=oxford_pets \
SEED=1 \
RES=96 \
RESET_LOREAL=1 \
bash scripts/run_lr_base_new.sh

Results and logs are written under:

$OUTPUT_ROOT/logs/CoOp_LOREAL/$DATASET/res$RES/seed$SEED/
$OUTPUT_ROOT/output/CoOp_LOREAL/base2new/train_base/$DATASET/

The summary file is:

$OUTPUT_ROOT/logs/CoOp_LOREAL/$DATASET/res$RES/seed$SEED/lr_base_new_summary.txt

Useful Environment Overrides

The example script exposes common experiment knobs through environment variables:

DATASET=oxford_flowers
SEED=2
RES=96
LOREAL_DIM=64
LOREAL_COEF1=0.5
LOREAL_COEF2=1.0
LOREAL_N_ATT=1
LOREAL_PROMPT_ORDER=ctx_attr_cls

Example:

DATASET=oxford_pets SEED=2 RESET_LOREAL=1 LOREAL_DIM=64 bash scripts/run_lr_base_new.sh

Acknowledgements

This codebase builds on:

• CoOp
• Dassl.pytorch
• CLIP
• DPC
• ATPrompt

Citation

@inproceedings{wang2026loreal,
  title={LOREAL: Mitigating Low-Resolution Challenges in Vision-Language Models with Attribute-driven Prompt Self-Distillation},
  author={Wang, Xucong and Wang, Pengkun and Zhao, Zhe and Yu, Liheng and Mao, Rui and Wang, Yang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={39152--39163},
  year={2026}
}