Google launches Gemini Omni 1.2 Flash with 15-second video generation

Extended video length and control

Google has officially released Gemini Omni 1.2 Flash, a model designed to make AI-powered video production more accessible for creators and businesses. As Pulse reported, one of the primary upgrades is the ability to generate native video clips up to 15 seconds long. This is an increase from the previous 10-second limit, allowing for more substantial content like product demonstrations or educational clips.

The model provides users with granular control by allowing them to specify video durations in one-second increments starting from three seconds. This flexibility is intended to help advertisers and content creators build specific types of short-form media. By offering these precise increments, Google aims to give users better predictability over the final output length.

Beyond just length, the model improves how it follows specific instructions. It can now use character images, product photography, and even short motion clips as references. These inputs help the AI maintain consistency in important details like facial features, typography, and brand colors across different shots.

Cost-effective production workflows

To help businesses manage costs, Google introduced a new workflow that prioritizes efficiency during the creative process. Users can now generate initial drafts at lower resolutions, specifically 360p or 720p. Once a draft is approved, they can upscale the final selected output to high-definition 1080p or 4K quality.

This approach reduces the heavy computing costs typically associated with generating high-resolution video repeatedly. Instead of starting from scratch every time a change is needed, the model supports conversational editing. This means users can ask for specific adjustments to the existing clip, such as changing the weather, lighting, or the background.

Because the underlying video sequence remains the same during these edits, the process is much faster than traditional generation. This capability is particularly useful for creating multiple variations of the same advertisement. It allows teams to test different environments or products while keeping the core action consistent.

Advanced motion and speech

Gemini Omni 1.2 Flash also introduces significant upgrades to how the AI handles movement and dialogue. It can now generate complex cinematic camera movements, such as first-person flythroughs or dramatic zoom-outs. These movements are designed to be more continuous and consistent, with fewer unwanted cuts between different visual scales.

The model also improves speech generation by creating more natural-sounding dialogue. It features better synchronization between the spoken words and the facial movements of the characters. Additionally, it can produce more distinct voices for multiple speakers, which is useful for creating professional productions with virtual presenters.

For motion specifically, the system can take a motion reference up to three seconds long. This helps the AI replicate specific movements between different subjects. Google demonstrated these capabilities using examples of architectural environments and animated sequences to show how the technology handles complex visual transfers.

Enterprise adoption and safety

Several major organizations are already testing the technology in their production pipelines. Adobe plans to include the model in Firefly to provide users with longer outputs and better image consistency. Other companies like Monks and Whering are evaluating the model for visual effects and accurately representing fabrics in virtual styling applications.

Google is supporting these large-scale users through Google Cloud infrastructure. This includes Provisioned Throughput, which gives organizations a predictable amount of computing capacity for their workloads. The model is also being integrated into existing tools like Google Vids, Flow, and Google Ads to reach a wider audience of creators.

To address safety and transparency, all generated outputs carry an imperceptible SynthID watermark. This technology is designed to identify the content as AI-generated without affecting the visual quality. While the model is powerful, Google noted that it still needs work on fast-moving sports footage and complex physical interactions.