OpenAI Enhances Sora 2 Model

- The original Sora model was first previewed in February 2024, with Sora 2 being released on September 30, 2025, as part of a dedicated mobile social app. - Beyond inpainting, Sora 2 introduced synchronized audio generation for dialogue and sound effects, and a "Cameo" feature that allows users to insert their own face and voice into the generated videos. - Video inpainting is a computer vision technique that restores or alters video by filling missing regions using spatial and temporal data from adjacent frames to ensure seamless visual coherence. This process is computationally intensive, especially in dynamic scenes with camera or object motion. - The model's real-time inference relies on powerful GPUs, such as NVIDIA's H100, which are designed for the parallel processing and large-scale matrix operations inherent in the transformer architectures used by generative AI. OpenAI has also partnered with chipmaker Cerebras to build out high-speed inference capacity. - Key competitors include Google's Veo 3, which is often cited for producing higher-quality and more realistic video, and ByteDance's Seedance 2.0, which supports higher resolution (2K vs. Sora's 1080p) and more complex multimodal inputs. - A significant challenge for generative video models, including Sora 2, is maintaining temporal consistency; this can result in visual artifacts like distorted limbs, objects warping, or unnatural physics over the duration of a clip. - The energy consumption required for AI video generation is substantial, with a study from Hugging Face warning that power demands grow exponentially with video length and resolution, posing a potential barrier to widespread on-device deployment and a challenge to climate goals. - While the first generation of Sora could create silent videos up to a minute long, Sora 2's clips are shorter, with a focus on higher physical accuracy, audio synchronization, and features tailored for a TikTok-style social media application.