Google Releases Gemini Omni 1.2 Flash for AI Video Generation
Google DeepMind has launched Gemini Omni 1.2 Flash, bringing 15-second native video generation, multi-turn conversational editing, and a 30% price cut.
TL;DR
- Google released Gemini Omni 1.2 Flash with 15-second native video generation, conversational editing, and 30% lower pricing [^1].
- The updated model improves reference adherence and visual consistency while adding imperceptible SynthID watermarking to generated outputs [^1].
Background
Generative video models allow developers and creators to produce synthetic video clips from text, image, or audio prompts. Historically, video generation models faced significant hurdles, including short clip durations, visual morphing across frames, inconsistent character details, and high computing costs. To make AI video practical for commercial workflows, providers aim to extend clip lengths, maintain consistent subjects across camera turns, and enable incremental editing without requiring complete re-generations.
What happened
Google DeepMind announced the general availability of Gemini Omni 1.2 Flash across the Gemini API, Google Cloud, and Google consumer applications [^1]. Announced by Product Manager Anish Nangia, the update introduces native video creation and editing for clips up to 15 seconds long, expanding from the 10-second limit in version 1.1 [^1]. Users can set video duration in one-second increments ranging from 3 to 15 seconds [^1].
The new release reduces generation prices by 30% compared to Gemini Omni 1.1, costing $0.07 per second for 720p video [^1]. To manage compute expenses, the workflow supports drafting initial clips at 360p or 720p, applying targeted follow-up edits, and upscaling finalized selections to 1080p or 4K [^1]. Multi-turn conversational editing enables users to alter backgrounds, lighting, weather, or aspect ratios across full 15-second clips without starting over from scratch [^1].
Gemini Omni 1.2 Flash enhances reference adherence by combining character or product photographs, audio cues, and short motion clips up to three seconds long [^1]. Google reported tighter subject consistency across camera transitions, smoother continuous camera movements like FPV flythroughs, and natural multi-speaker speech with lip-syncing [^1]. Outputs generated by the model carry an imperceptible SynthID watermark [^1].
Why it matters
Expanding video generation to 15 seconds while offering multi-turn conversational editing addresses major friction points in generative media workflows [^1]. Previously, modifying minor scene elements required re-rendering entire video files, multiplying compute costs and introducing unexpected visual alterations. Enabling turn-by-turn adjustments while retaining character and product details gives creators finer control over final assets [^1].
Additionally, lowering generation prices by 30% and introducing low-resolution drafting options lowers financial barriers for enterprise video production at scale [^1]. As AI video tools expand across developer APIs and commercial platforms, embedding provenance signals like SynthID watermarks becomes critical for tracking synthetic media [^1].
Practical example
Say Elena manages digital campaigns for a clothing retail brand. Her team needs to create a series of 15-second promotional video ads showcasing a new jacket across diverse settings [^1].
Using Gemini Omni 1.2 Flash through the Gemini API, Elena inputs a model photograph alongside front-and-back shots of the outfit [^1]. She generates a 15-second virtual try-on clip drafted at 720p for $0.07 per second [^1]. The model maintains the subject's identity and clothing drape across a 360-degree turn [^1].
Instead of generating a new scene from scratch to test alternate environments, Elena uses multi-turn conversational prompts to swap the background from a pavilion tent to an oak forest, a market fair, and a stone castle while preserving the character's performance and camera framing [^1]. Satisfied with the variants, she upscales the final renders to 1080p with embedded SynthID watermarks [^1].