.svg){kind=spacer}
Details
- Google DeepMind introduces Veo 3.1, the latest version of its text-to-video model, now available to Flow by Google users and through the Gemini API for developers.
- The update significantly advances narrative comprehension, enabling the model to maintain consistent characters, lighting, and photorealistic textures in longer video sequences.
- A new "ingredients to video" mode allows creators to upload multiple reference images, blending elements from each into a unified, sound-enabled scene.
- The "scene extension" feature auto-generates continuing footage of 60 seconds or more, linking clips by leveraging the previous clip's final moments for seamless continuity.
- The "first-and-last-frame" capability generates smooth transitions and camera movements, automatically filling in between endpoints without the need for manual key-framing.
- Parallel audio track generation syncs sound precisely to onscreen action, cutting down on post-production sound design efforts.
- The improvements stem from enhanced video-language pre-training and a reinforced diffusion decoder, though Veo's model weights are not open source.
Impact
Veo 3.1 directly takes on OpenAI's Sora in the long-form generative video space, surpassing the previous standard for minute-long, story-consistent outputs. Its integration with Gemini API positions Google DeepMind as a strong multimodal platform contender, pushing competitors to innovate in both video and audio generation. As AI-generated video becomes more sophisticated, it could reshape content creation workflows and prompt new regulatory debates around intellectual property.
