Computer vision solutions · 10–15%

Image & video generation

~2 min read

Image generation

An image-generation model creates original images from a text description — it generates, it doesn’t search a catalog.

• Models: gpt-image-1 (OpenAI), FLUX (Black Forest Labs). Catalog filter: inference task = text to image.
• Call client.images.generate(model, prompt, n, size). The response carries the image as base64 in b64_json; decode it and save it to a file.
• Generated images are original, not retrieved.
• Editing workflows: inpainting — regenerating only a masked region of an image while leaving the rest untouched; plus prompt-driven modifications and reference images to steer output.

Video generation — Sora 2

Sora 2 is OpenAI’s video-generation model in Foundry: it creates video (with audio) from a text prompt, a reference image, or by remixing an existing video.

Async pattern — generation takes minutes, so the API works in three steps:

1. Start the job with videos.create().
2. Poll videos.retrieve(id) until the status leaves queued/in_progress.
3. Download the result with videos.download_content(id, variant="video").

Possible states: queued, in_progress, completed, failed, cancelled.

Prompting videos

Describe the shot like a storyboard, in this order: camera framing, then subject detail, then the action in beats, then lighting and palette, then style. Specific beats (“takes four steps, pauses”) beat vague verbs. Shorter clips follow instructions better — stitch two 4s clips over one 8s.