OpenAI Sora 2

Über OpenAI

Gegründet 2015 · San Francisco, USA

OpenAI was founded in December 2015 by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, Wojciech Zaremba, John Schulman, Andrej Karpathy and others. The company shipped Sora 1 in December 2024 inside ChatGPT and the standalone Sora app. Sora 2 launched in 2025 as the substantially upgraded successor: longer durations, much better physics, native synchronized audio (sound effects, music and dialogue) and an entirely new social product called the Sora app, which features a TikTok-like feed of user-generated AI clips and the 'cameo' feature that lets users insert their own and friends' likenesses into generated videos. Sora 2 cemented OpenAI's position alongside Google DeepMind's Veo 3 family as the frontier of consumer-accessible video generation in 2025.

Architektur

Diffusion Transformer on spacetime patches with joint audio-video diffusion

Sora 2 builds on the Sora 1 'spacetime patches' DiT architecture: a 3D causal VAE encodes video into a spatio-temporal latent and a transformer denoiser is trained on patches drawn from that latent. The headline architectural change is a joint audio-video diffusion pipeline that produces synchronized soundtrack -- including sound effects, ambient noise, music swells and short dialogue with lip-sync -- alongside the visual track. Sora 2 also adds an explicit 'cameo' conditioning channel that locks character identity to a user-provided reference video / images while enforcing OpenAI's identity-consent and impersonation policies. Native generation runs up to ~25-60 seconds depending on tier, at up to 1080p. Conditioning uses a GPT-family text encoder with dense recaptioning, plus optional image and video references. Public benchmark results and OpenAI's own internal evaluations report substantial gains over Sora 1 in motion fidelity, physical plausibility and prompt adherence.

Parameter

Undisclosed

Funktionen

• Native synchronized audio (sound effects, ambient sound, dialogue with lip-sync)
• Up to ~25-60 second clips at up to 1080p depending on tier
• 'Cameo' feature: insert user / friend likeness with consent verification
• Strong physical plausibility (water, fabric, crowds, gravity)
• Text-to-video, image-to-video, video-to-video, in-painting and extension
• Social Sora app with TikTok-like feed of AI clips
• Dense GPT-class recaptioning pipeline for strong prompt adherence
• Available via Sora app and ChatGPT Plus / Pro; OpenAI shut down its video API (Videos API) on 24 September 2026
• Best for: cinematic shorts with sound, social-media content, character cameos.

Training & Lizenz

Massive curated multilingual audio-video corpus combining licensed footage, web data and partner sources, with dense synthetic audio-visual captions; exact size undisclosed.

Lizenz: Proprietary commercial licence via OpenAI terms; commercial use on paid plans subject to content policy; mandatory C2PA metadata and visible watermark on outputs; cameo feature gated by identity-verification flow.

Sicherheitstests: Extensive external red-team, cameo identity-consent enforcement, public-figure and CSAM filters, deepfake / impersonation guards; published system card.