OpenAI¶

OpenAI provides image generation via DALL-E and GPT Image, and video generation via Sora. Tarash uses the official openai Python SDK.

Installation¶

pip install tarash-gateway[openai]

Image Generation¶

Quick Example¶

from tarash.tarash_gateway import generate_image
from tarash.tarash_gateway.models import ImageGenerationConfig, ImageGenerationRequest

config = ImageGenerationConfig(
    provider="openai",
    model="dall-e-3",
    api_key="sk-...",
)
request = ImageGenerationRequest(
    prompt="A futuristic cityscape at night, photorealistic",
    size="1792x1024",
    quality="hd",
    style="vivid",
    n=1,
)
response = generate_image(config, request)
print(response.images[0])

Parameters¶

Supported Models¶

Video Generation¶

Quick Example¶

from tarash.tarash_gateway import generate_video
from tarash.tarash_gateway.models import VideoGenerationConfig, VideoGenerationRequest

config = VideoGenerationConfig(
    provider="openai",
    model="openai/sora-2",
    api_key="sk-...",
)
request = VideoGenerationRequest(
    prompt="A serene mountain lake at golden hour",
    duration_seconds=8,
    aspect_ratio="16:9",
)
response = generate_video(config, request)
print(response.video)

Parameters¶

Supported Models¶

Image-to-Video¶

Sora supports using a reference image as the first frame of the generated video. Pass a single image via image_list — only the first image is used regardless of its type value.

from tarash.tarash_gateway.models import ImageType, VideoGenerationRequest

request = VideoGenerationRequest(
    prompt="The car drives away into the sunset",
    image_list=[ImageType(image="https://example.com/car.jpg", type="reference")],
    duration_seconds=8,
    aspect_ratio="16:9",
)

Video Remix¶

To remix an existing Sora video, pass the original video_id via extra_params. Sora will use the existing video as a base and apply your prompt as a variation.

request = VideoGenerationRequest(
    prompt="Same scene but with snow",
    extra_params={"video_id": "video_abc123"},
)

Azure OpenAI¶

Azure OpenAI runs the same Sora and image models through Microsoft Azure. Use provider="azure_openai" and provide two additional required fields: base_url (your Azure resource endpoint) and api_version.

Quick Example¶

config = VideoGenerationConfig(
    provider="azure_openai",
    model="my-sora-deployment",   # your Azure deployment name
    api_key="YOUR_AZURE_KEY",
    base_url="https://my-resource.openai.azure.com/",
    api_version="2024-05-01-preview",
)

Parameters, image/video generation, Image-to-Video, and Video Remix all work identically to the standard OpenAI provider above.

Azure-Specific Notes¶

Deployment names: Azure requires a custom deployment name instead of a model ID (e.g. my-sora-deployment rather than openai/sora-2).

api_version from URL: If base_url already contains ?api-version=..., it is extracted automatically.

Authentication: Use api_key for an Azure resource key, or set it to a bearer token obtained via azure-identity for Azure AD authentication.

Provider-Specific Notes¶

Client caching: Both sync and async OpenAI clients are cached per (api_key, base_url) pair. This is safe because OpenAI's clients support reuse across calls.

Content download: After generation, Tarash automatically downloads the video bytes so the response includes the content directly, not just a URL.