Idefics3-Llama3

Transformers version: >4.46.

Idefics3

Idefics3 is an open multimodal model that accepts arbitrary sequences of image and text inputs and produces text outputs. The model can answer questions about images, describe visual content, create stories grounded on multiple images, or simply behave as a pure language model without visual inputs. It improves upon Idefics1 and Idefics2, significantly enhancing capabilities around OCR, document understanding and visual reasoning.

We release the checkpoints under the Apache 2.0.

Model Summary

• Developed by: Hugging Face
• Model type: Multi-modal model (image+text)
• Language(s) (NLP): en
• License: Apache 2.0
• Parent Models: google/siglip-so400m-patch14-384 and meta-llama/Meta-Llama-3.1-8B-Instruct
• Resources for more information:
  • Idefics1 paper: OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents
  • Idefics2 paper: What matters when building vision-language models?
  • Idefics3 paper: Building and better understanding vision-language models: insights and future directions

Uses

Idefics3-8B can be used to perform inference on multimodal (image + text) tasks in which the input is composed of a text query along with one (or multiple) image(s). Text and images can be arbitrarily interleaved. That includes image captioning, visual question answering, etc. These model does not support image generation.

The post-training of Idefics3-8B involves only a supervised fine-tuning stage, without RLHF alignment. As a result, the model may produce short answers or require prompt iterations to fully address the user's request. Adding a prefix to the assistant's response, such as "Let's fix this step by step" has been found to effectively influence the generated output.

To fine-tune Idefics3-8B on a specific task, we provide a fine-tuning tutorial. Other resources for the fine-tuning of Idefics2 (can easily be adapted to Idefics3):

• With the TRL library: Script

Technical summary

Idefics3 demonstrates a great improvement over Idefics2, especially in document understanding tasks. It serves as a strong foundation for various use-case specific fine-tunings.

Idefics3 introduces several changes compared to Idefics2:

• We use 169 visual tokens to encode a image of size 364x364. Each image is divided into several sub images of sizes at most 364x364, which are then encoded separately.
• For the fine-tuning datasets, we have extended The Cauldron and added several datasets, including Docmatix. We will push soon these datasets to the same repo of The Cauldron (TODO).

More details about the training of the model is available in our technical report.

How to Get Started

This section shows snippets of code for generation for Idefics3-8B.

import requests
import torch
from PIL import Image
from io import BytesIO

from transformers import AutoProcessor, AutoModelForVision2Seq
from transformers.image_utils import load_image

DEVICE = "cuda:0"

# Note that passing the image urls (instead of the actual pil images) to the processor is also possible
image1 = load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg")
image2 = load_image("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg")
image3 = load_image("https://cdn.britannica.com/68/170868-050-8DDE8263/Golden-Gate-Bridge-San-Francisco.jpg")

processor = AutoProcessor.from_pretrained("HuggingFaceM4/Idefics3-8B-Llama3")
model = AutoModelForVision2Seq.from_pretrained(
    "HuggingFaceM4/Idefics3-8B-Llama3", torch_dtype=torch.bfloat16
).to(DEVICE)

# Create inputs
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "What do we see in this image?"},
        ]
    },
    {
        "role": "assistant",
        "content": [
            {"type": "text", "text": "In this image, we can see the city of New York, and more specifically the Statue of Liberty."},
        ]
    },
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "And how about this image?"},
        ]
    },
]
prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(text=prompt, images=[image1, image2], return_tensors="pt")
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}

# Generate
generated_ids = model.generate(**inputs, max_new_tokens=500)
generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)

print(generated_texts)

Text generation inference

TODO.

Model optimizations

If your GPU allows, we first recommend loading (and running inference) in half precision (torch.float16 or torch.bfloat16).

model = Auto