Image-Text-to-Text
PEFT
Safetensors
Metric-ViPer / README.md
miaw1419's picture
Update README.md
359509e verified
|
raw
history blame
1.68 kB
metadata
{}

ViPer: Metric

GitHub: https://github.com/sogandstorme/ViPer_Personalization

Example

To use ViPer, start by cloning it from our GitHub page. For effective personalization, we recommend commenting on at least eight images.

git clone https://github.com/sogandstorme/ViPer_Personalization.git
cd ViPer_Personalization
from ViPer import (
    set_device,
    load_images,
    initialize_processor_and_model,
    prepare_prompt_and_inputs,
    generate_texts,
    extract_features,
    initialize_pipelines,
    generate_images
)

# Ensure that the order of the comments matches the path of the images they refer to.

comments = [
    "These are beautiful, intricate patterns. Very elegant, and the teal blue colors are lovely. I love the flowing lines.",
    "The colors here don't quite work for me. They feel a bit unmatched and artificial. The concept also seems a bit boring and artificial to me.",
    ]

image_paths = [
    "/images/6.png",
    "/images/9.png"
]

prompts = [
    "Whimsical tea party in a bioluminescent forest",
    "Tiny houses on top of each other above clouds"
]

output_dir = "results/"

device = set_device("cuda:0")
    
# Initialize processor, model and inputs
images = load_images(image_paths)
processor, model = initialize_processor_and_model(device)
inputs = prepare_prompt_and_inputs(processor, images, comments)

# Generate and extract vp
generated_texts = generate_texts(processor, model, inputs)
vp_pos, vp_neg = extract_features(generated_texts)

# Initialize pipelines and generate images
pipe, refiner = initialize_pipelines(device)
generate_images(pipe, refiner, prompts, vp_pos, vp_neg, output_dir)