The authors initialize the weights of the image encoder and large language model from pre-trained checkpoints and keep them frozen while training the Querying Transformer, which is a BERT-like Transformer encoder that maps a set of "query tokens" to query embeddings, which bridge the gap between the embedding space of the image encoder and the large language model.

The goal for the model is simply to predict the next text token, giving the query embeddings and the previous text.

This allows the model to be used for tasks like:

image captioning
visual question answering (VQA)
chat-like conversations by feeding the image and the previous conversation as prompt to the model

Direct Use and Downstream Use

You can use the raw model for conditional text generation given an image and optional text. See the model hub to look for fine-tuned versions on a task that interests you.

Bias, Risks, Limitations, and Ethical Considerations

BLIP2-FlanT5 uses off-the-shelf Flan-T5 as the language model. It inherits the same risks and limitations from Flan-T5:

Language models, including Flan-T5, can potentially be used for language generation in a harmful way, according to Rae et al. (2021). Flan-T5 should not be used directly in any application, without a prior assessment of safety and fairness concerns specific to the application.

BLIP2 is fine-tuned on image-text datasets (e.g. LAION ) collected from the internet. As a result the model itself is potentially vulnerable to generating equivalently inappropriate content or replicating inherent biases in the underlying data.

BLIP2 has not been tested in real world applications. It should not be directly deployed in any applications. Researchers should first carefully assess the safety and fairness of the model in relation to the specific context they’re being deployed within.

How to use

For code examples, we refer to the documentation, or refer to the snippets below depending on your usecase:

Running the model on CPU

Click to expand

python

import requests
from PIL import Image
from transformers import BlipProcessor, Blip2ForConditionalGeneration

processor = BlipProcessor.from_pretrained("Salesforce/blip2-flan-t5-xxl")
model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-flan-t5-xxl")

img_url = 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/demo.jpg' 
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert('RGB')

question = "how many dogs are in the picture?"
inputs = processor(raw_image, question, return_tensors="pt")

out = model.generate(**inputs)
print(processor.decode(out[0], skip_special_tokens=True))

Running the model on GPU

In full precision

Click to expand

python

# pip install accelerate
import requests
from PIL import Image
from transformers import Blip2Processor, Blip2ForConditionalGeneration

processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xxl")
model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-flan-t5-xxl", device_map="auto")

img_url = 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/demo.jpg' 
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert('RGB')

question = "how many dogs are in the picture?"
inputs = processor(raw_image, question, return_tensors="pt").to("cuda")

out = model.generate(**inputs)
print(processor.decode(out[0], skip_special_tokens=True))

In half precision (`float16`)

Click to expand

python

# pip install accelerate
import torch
import requests
from PIL import Image
from transformers import Blip2Processor, Blip2ForConditionalGeneration

processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xxl")
model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-flan-t5-xxl", torch_dtype=torch.float16, device_map="auto")

img_url = 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/demo.jpg' 
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert('RGB')

question = "how many dogs are in the picture?"
inputs = processor(raw_image, question, return_tensors="pt").to("cuda", torch.float16)

out = model.generate(**inputs)
print(processor.decode(out[0], skip_special_tokens=True))

In 8-bit precision (`int8`)

Click to expand

python

# pip install accelerate bitsandbytes
import torch
import requests
from PIL import Image
from transformers import Blip2Processor, Blip2ForConditionalGeneration

processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xxl")
model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-flan-t5-xxl", load_in_8bit=True, device_map="auto")

img_url = 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/demo.jpg' 
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert('RGB')

question = "how many dogs are in the picture?"
inputs = processor(raw_image, question, return_tensors="pt").to("cuda", torch.float16)

out = model.generate(**inputs)
print(processor.decode(out[0], skip_special_tokens=True))

LanguageMachines/blip2-flan-t5-xxl

BLIP-2, Flan T5-xxl, pre-trained only

Model description

Direct Use and Downstream Use

Bias, Risks, Limitations, and Ethical Considerations

How to use

Running the model on CPU

Running the model on GPU

In full precision

In half precision (float16)

In 8-bit precision (int8)

LanguageMachines/
blip2-flan-t5-xxl

In half precision (`float16`)

In 8-bit precision (`int8`)