Image Generations

Some of the models can provide image generation capabilities. You can provide some descriptions of the image you would like to generate, and let the model generate one or multiple pictures in the output.

If you're used to interacting with the chat/image-understanding models, the image generation is a bit different from them. You only need to send a prompt text in the request, instead of a list of messages with system/user/assistant roles. When you sent the prompt for image generation, your prompt will be revised by a chat model, and then sent to the image generation model.

• n: Number of image(s) to generate (1-10, default to 1)
• response_format: "url" or "b64_json". If "url" is specified, the response will return a url to the image(s) in data[index].url; if "b64_json" is specified, the response will return the image(s) in base64 encoded format in data[index].b64_json.

Note: quality, size or style are not supported by xAI API at the moment.

The image generation is offered at a different endpoint https://api.x.ai/v1/images/generations from the chat and image-understanding models that share https://api.x.ai/v1/chat/completions. The endpoint is compatible with OpenAI SDK, so you can keep using the same base_url of https://api.x.ai/v1. Note: Anthropic SDK is not supported for image generation and is deprecated for all endpoints.

You can set "model": "grok-2-image" in the request body to use the model. The generated image will be in jpg format.

import os

from xai_sdk import Client

client = Client(api_key=os.getenv('XAI_API_KEY'))

response = client.image.sample(
    model="grok-2-image",
    prompt="A cat in a tree",
    image_format="url"
)

print(response.url)

The Python and JavaScript examples will print out url of the image on xAI managed storage.

This is an example image generated from the above prompt:

Instead of getting an image url by default, you can choose to get a base64 encoded image instead. To do so, you need to specify the response_format parameter to "b64_json".

import os

from xai_sdk import Client

client = Client(api_key=os.getenv('XAI_API_KEY'))

response = client.image.sample(
    model="grok-2-image",
    prompt="A cat in a tree",
    image_format="base64"
)

print(response.image) # returns the raw image bytes

You will get a b64_json field instead of url in the response image object.

You can generate up to 10 images in one request by adding a parameter n in your request body. For example, to generate four images:

import os

from xai_sdk import Client

client = Client(api_key=os.getenv('XAI_API_KEY'))

response = client.image.sample_batch(
    model="grok-2-image",
    prompt="A cat in a tree",
    n=4
    image_format="url",
)

for image in response:
    print(response.url)

If you inspect the response object, you can see something similar to this:

{
  "data": [
    {
      "b64_json": "data:image/png;base64,...",
      "revised_prompt": "..."
    }
  ]
}

Before sending the prompt to the image generation model, the prompt will be revised by a chat model. The revised prompt from chat model will be used by image generation model to create the image, and returned in revised_prompt to the user.

To see the revised prompt with SDK:

# ... Steps to make image generation request

print(response.prompt)

For example: