LIGER

Code for the paper Long-horizon Visual Instruction Generation with Logic and Attribute Self-reflection (ICLR 2025)

Preparation

Prepare the following projects according to their instructions:

• Lama

• LISA

• SD-Inpaint

• DragonDiffusion

Then install the requirements by running:

pip install -r requirements.txt

Prepare your Openai API key by:

export OPENAI_API_KEY='Your_API_Key'

Proposed Dataset

The proposed dataset contains 569 long-horizon tasks are in the step_data folder.

The GT_description subfolder contains the annotated descriptions for each step in every task.

The all_tasks_with_logic subfolder contains the GPT generated task step descriptions, historical prompt, and logic step annotations.

Running

Run following command for visual instruction generation:

python run.py

Evaluation

Run following command for evaluation:

python metric_clip_and_text.py    #CLIP-score and BERT-score evaluation
python metric_dino.py             #DINO-score evaluation
python metric_GPT.py              #GPT evaluation