Gemini Omni 1.2 Flash adds 15-second clips and multi-turn editing at lower cost
Google · Anish NangiaProduct Manager, Google DeepMind
- Models
- Developer tools
- Industry
- #gemini
- #video generation
- #google deepmind
- #synthid
Google DeepMind announced Gemini Omni 1.2 Flash, a generally available video generation model on the Gemini API and Google Cloud. The publisher says it raises native clip length from 10 to 15 seconds, improves reference adherence, adds multi-turn conversational editing, and cuts price 30% to $0.07 per second for 720p. Outputs carry an imperceptible SynthID watermark, and Google notes development continues on fast action, rapid script changes and complex physics.
What Google announced
Google DeepMind is rolling out Gemini Omni 1.2 Flash, described as generally available across the Gemini API, Google Cloud and Gemini for business. The mechanism the post describes is a single model that generates and edits video natively. Clip length rises from 10 to 15 seconds, with duration selectable in one-second increments between 3 and 15. Editing is conversational: rather than re-running a whole generation to change a background, lighting, product or aspect ratio, a user issues follow-up instructions across turns on the same clip. Reference adherence is the other stated improvement, with character, product or style images, motion clips up to three seconds, and audio cues said to hold detail such as faces, garment cut and packaging typography across camera moves. Google says the price falls 30% versus Omni 1.1 to $0.07 per second at 720p, and that a draft-fast-then-upscale workflow (360p or 720p first, 1080p or 4K later) keeps costs down. Outputs carry an imperceptible SynthID watermark. The post shows comparison examples claiming less drift and fewer jump cuts, but these are publisher-provided demonstrations, not independent evaluations. Google also states that work continues on fast sports action, rapid script changes and complex physics.
Where a creative producer could start
A social creative producer at a mid-sized brand could test Omni 1.2 Flash on ad variants rather than hero footage. The post's claimed fit is exactly this: one reference image set, a 15-second spot, then multi-turn edits that swap background or aspect ratio for different placements. A producer might take one existing product-photo library, generate a short virtual try-on or catalog-spin sequence, and ask for background variants across turns, then judge whether identity, garment drape and brand colors survive the edits well enough to replace manual re-briefing. This is a plausible trial, not a proven result; the comparison claims come only from Google's examples.
How to run a defensible trial
A team could try a small controlled comparison: pick five reference sets (say, one face, one product, one style frame) and run each prompt on both Omni 1.1 and 1.2 Flash, holding prompts and reference images fixed. Baseline: the current manual or prior-model workflow, including time spent and number of regenerations per accepted clip. Measure how many first outputs pass a reviewer's reference-adherence check, how many turn-by-turn edits stay on-model, and total cost per approved clip using the published per-second rate. Failure checks should include deliberate stress cases: fast motion, rapid dialogue changes, hands and text, and physics-heavy scenes the post flags as still improving. A human creative or brand reviewer should approve every clip, and teams should confirm that the SynthID watermark and any disclosure obligations fit their channel rules before publishing.
Trade-offs and when to wait
The lower price and conversational edits shift cost from regeneration to iteration, but they also mean review effort moves to checking each turn for drift, which the post does not quantify. Prerequisites include API or Google Cloud access, reference assets of suitable quality, and a pipeline that can upscale from 720p. Unknowns: independent quality benchmarks, how consistency holds on long multi-turn sequences, latency, rate limits and throughput beyond the mention of Provisioned Throughput. A team should not use this where the post itself flags weakness, such as fast sports action or complex physics, and should avoid treating a 15-second native cap as a substitute for longer narrative editing. Where legal or brand rules require provenance tracking, verify watermark and metadata handling first.
Context and limitations
- Based on a short source excerpt; the full article was not analyzed.
- All capability, price and consistency claims come from Google's announcement; no independent testing is reported.
- Google itself says fast sports action, rapid script changes and complex physics still need improvement.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.