LiAutoAD/Ristretto-3B

Introduction

We are excited to introduce Ristretto, our newest Vision language model (VLM) that represents a significant step forward in the field. Ristretto features a capability to deploy dynamic image tokens, enables flexible adjustment of image token quantities based on task requirements while enhancing the projector architecture to support dynamic token configurations. This new model delivers improved performance and versatility compared to its predecessors through its refined architecture and advanced training approach.

Key Innovations

Coming soon...

Environment Setup

pip install torch>=2.3.0
pip install transformers==4.37.0

How to use?

import torch
import torchvision.transforms as T
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
import requests
from io import BytesIO

IMAGENET_MEAN = (0.5, 0.5, 0.5)
IMAGENET_STD = (0.5, 0.5, 0.5)

def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform

def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio

def dynamic_preprocess(image, min_num=1, max_num=10, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

def load_image(image_data, input_size=384, max_num=10):
    image = Image.open(image_data).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

model_path = 'LiAutoAD/Ristretto-3B'
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)



image_url = 'https://github.com/user-attachments/assets/83258e94-5d61-48ef-a87f-80dd9d895524'
response = requests.get(image_url)
image_data = BytesIO(response.content)
pixel_values = load_image(image_data, max_num=10).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens=1024, do_sample=True)

# The recommended range for `num_image_token` is 64 to 576, and the value can be adjusted based on task requirements.
num_image_token = 256

# pure-text conversation
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}         Assistant: {response}')

# text-image conversation && multi-round conversation
question = '<image>         Please describe the image.'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}         Assistant: {response}')


question = 'What is best title for the image?'
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
print(f'User: {question}         Assistant: {response}')

Evaluation

Benchmark	Qwen2.5-VL-3B	InternVL2.5-4B	Ristretto-3B
MMBench-TEST-avg	76.8	78.2	80.1
MMStar	56.3	58.7	62.6
MMMU-VAL	51.2	51.8	49.1
MathVista-MINI-test	61.2	60.8	67.9
HallucinationBench	46.6	46.6	50.2
AI2D	81.4	81.4	84.3
OCRBench	82.8	82.0	84.0
MMVet	60.0	61.5	61.8
Average	64.5	65.1	67.6

We use VLMEvalKit to evaluate Ristretto-3B. Other results are taken from OpenCompass

License Agreement

All of our open-source models are licensed under the Apache-2.0 license.

Citation