seed_llama_templates.py

# from SEED-LLaMA Paper 
# Video 관련 태스크는 생략

BOI_TOKEN = '<img>'
EOI_TOKEN = '</img>'
IMG_TOKEN = '<img_{:05d}>'


# .format(caption)
text_to_image_generation = "USER: {} Please generate an image. \nASSISTANT: " 

# .format(image1, instruction) - 보통 Image Editing
multimodal_prompt_image_generation = "USER: {} {} Please generate an image.\nASSISTANT: "

# .format(image, question)
image_conversation = "USER: {} {}\nASSISTANT: "

# .format(image1, image2, question)
multi_image_understanding = "USER: This is the first image. {} This is the second image. {} {}\nASSISTANT: "

# .format(image)
image_captioning = "USER: Please provide an accurate and concise description of the given image.\nASSISTANT: "

# .format(image, question)
image_qa = "USER: {} {} Please provide an accurate answer consisting of only one word or phase.\nASSISTANT: "