Use Visual Question Answering (VQA) to get image information

Visual Question Answering (VQA) lets you provide an image to the model and ask a question about the image's contents. In response to your question you get one or more natural language answers.

Sample VQA image, question and answers in the console
Image source (shown in Google Cloud console): Sharon Pittaway on Unsplash
Prompt question: What objects are in the image?
Answer 1: marbles
Answer 2: glass marbles

Languages supported

VQA is available in the following languages:

  • English (en)

Performance and limitations

The following limits apply when you use this model:

Limits Value
Maximum number of API requests (short-form) per minute per project 500
Maximum number of tokens returned in response (short-form) 64 tokens
Maximum number of tokens accepted in request (VQA short-form only) 80 tokens

The following service latency estimates apply when you use this model. These values are meant to be illustrative and are not a promise of service:

Latency Value
API requests (short-form) 1.5 seconds

Locations

A location is a region you can specify in a request to control where data is stored at rest. For a list of available regions, see Generative AI on Vertex AI locations.

Responsible AI safety filtering

The image captioning and Visual Question Answering (VQA) feature model doesn't support user-configurable safety filters. However, the overall Imagen safety filtering occurs on the following data:

  • User input
  • Model output

As a result, your output may differ from the sample output if Imagen applies these safety filters. Consider the following examples.

Filtered input

If the input is filtered, the response is similar to the following:

{
  "error": {
    "code": 400,
    "message": "Media reasoning failed with the following error: The response is blocked, as it may violate our policies. If you believe this is an error, please send feedback to your account team. Error Code: 63429089, 72817394",
    "status": "INVALID_ARGUMENT",
    "details": [
      {
        "@type": "type.googleapis.com/google.rpc.DebugInfo",
        "detail": "[ORIGINAL ERROR] generic::invalid_argument: Media reasoning failed with the following error: The response is blocked, as it may violate our policies. If you believe this is an error, please send feedback to your account team. Error Code: 63429089, 72817394 [google.rpc.error_details_ext] { message: \"Media reasoning failed with the following error: The response is blocked, as it may violate our policies. If you believe this is an error, please send feedback to your account team. Error Code: 63429089, 72817394\" }"
      }
    ]
  }
}

Filtered output

If the number of responses returned is less than the sample count you specify, this means the missing responses are filtered by Responsible AI. For example, the following is a response to a request with "sampleCount": 2, but one of the responses is filtered out:

{
  "predictions": [
    "cappuccino"
  ]
}

If all the output is filtered, the response is an empty object similar to the following:

{}

Use VQA on an image (short-form responses)

Use the following samples to ask a question and get an answer about an image.

REST

For more information about imagetext model requests, see the imagetext model API reference.

Before using any of the request data, make the following replacements:

  • PROJECT_ID: Your Google Cloud project ID.
  • LOCATION: Your project's region. For example, us-central1, europe-west2, or asia-northeast3. For a list of available regions, see Generative AI on Vertex AI locations.
  • VQA_PROMPT: The question you want to get answered about your image.
    • What color is this shoe?
    • What type of sleeves are on the shirt?
  • B64_IMAGE: The image to get captions for. The image must be specified as a base64-encoded byte string. Size limit: 10 MB.
  • RESPONSE_COUNT: The number of answers you want to generate. Accepted integer values: 1-3.

HTTP method and URL: 

POST https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/imagetext:predict

Request JSON body: 

{
  "instances": [
    {
      "prompt": "VQA_PROMPT",
      "image": {
          "bytesBase64Encoded": "B64_IMAGE"
      }
    }
  ],
  "parameters": {
    "sampleCount": RESPONSE_COUNT
  }
}

To send your request, choose one of these options:

curl

Save the request body in a file named request.json, and execute the following command: 

curl -X POST \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     -H "Content-Type: application/json; charset=utf-8" \
     -d @request.json \
     "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/imagetext:predict"

PowerShell

Save the request body in a file named request.json, and execute the following command: 

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method POST `
    -Headers $headers `
    -ContentType: "application/json; charset=utf-8" `
    -InFile request.json `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models/imagetext:predict" | Select-Object -Expand Content
The following sample responses are for a request with "sampleCount": 2 and "prompt": "What is this?". The response returns two prediction string answers. 
{
  "predictions": [
    "cappuccino",
    "coffee"
  ]
}

Python

Before trying this sample, follow the Python setup instructions in the Vertex AI quickstart using client libraries. For more information, see the Vertex AI Python API reference documentation.

To authenticate to Vertex AI, set up Application Default Credentials. For more information, see Set up authentication for a local development environment.

In this sample you use the load_from_file method to reference a local file as the base Image to get information about. After you specify the base image, you use the ask_question method on the ImageTextModel and print the answers.


import vertexai
from vertexai.preview.vision_models import Image, ImageTextModel

# TODO(developer): Update and un-comment below lines
# project_id = "PROJECT_ID"
# input_file = "my-input.png"
# question = "" # The question about the contents of the image.

vertexai.init(project=project_id, location="us-central1")

model = ImageTextModel.from_pretrained("imagetext@001")
source_img = Image.load_from_file(location=input_file)

answers = model.ask_question(
    image=source_img,
    question=question,
    # Optional parameters
    number_of_results=1,
)

print(answers)

Use parameters for VQA

When you get VQA responses there are several parameters you can set depending on your use case.

Number of results

Use the number of results parameter to limit the amount of responses returned for each request you send. For more information, see the imagetext (VQA) model API reference.

Seed number

A number you add to a request to make generated responses deterministic. Adding a seed number with your request is a way to assure you get the same prediction (responses) each time. However, the answers aren't necessarily returned in the same order. For more information, see the imagetext (VQA) model API reference.

What's next