Video generation with Sora 2 in Microsoft Foundry

Video generation works differently from image generation: it runs as a long job that you start, poll and then download.

What Sora 2 does

• Sora 2 (sora-2, preview) is an Azure OpenAI model deployed in a Foundry project. It supports text to video, image to video and (generated) video to video, generates audio in the output, and supports remix.
• Billing is per second of video.
• Try it in the video playground (preview), which can set aspect ratio and duration and shows code.
• Generated videos are kept for 24 hours, so download them.

The asynchronous job pattern

• Sora 2 uses the v1 API (https://<resource>.openai.azure.com/openai/v1/) with the OpenAI client.
• Call client.videos.create(...), then poll GET /videos/{video_id} (client.videos.retrieve) until the status leaves queued/in_progress and becomes completed or failed.
• Then call download_content(video_id, variant="video") for the MP4. create_and_poll wraps this in async code.
• A typical job takes 1 to 5 minutes. The five endpoints are create, get status, download, list and delete.

Common trap: expecting video generation to be a single synchronous call that returns the file - you create a job, poll its status, then download the result.

Sora 2 parameters

Common trap: assuming input_reference is resized for you - a reference image that doesn’t match the requested size is rejected.

• Remix keeps the original’s framework, scene transitions and layout. Limit each remix to one clearly stated change for the best fidelity.

The original Sora model

• The original sora model (2025-05-02) used the jobs REST API (/video/generations/jobs) with width, height, n_seconds (1-20 s) and n_variants.
• Up to four variants were allowed, two at 720p and none at 1080p.
• Nine resolutions from 480x480 to 1920x1080 were supported, two jobs could run at once, and jobs expired after 24 hours.

Restrictions and limitations

• Content must be suitable for under-18 audiences.
• Copyrighted characters and music, real people (including public figures) and input images containing human faces are rejected. Sora 2 blocks IP and photorealistic content.
• A job that trips a filter ends failed with a failure_reason.
• The model can struggle with complex physics, cause and effect, left and right, and precise camera timing.
• Prompts work best in English or other Latin-script languages.

Get the whole book

This note is one section of Ultra Transcenders AI-103: Developing AI Apps and Agents on Azure, an independent study guide that explains every topic the exam covers by technology, with comparison tables, diagrams and the common traps, plus a glossary linked to Microsoft Learn.

About the book · Free AI-103 glossary · All AI-103 study notes