Gemini Omni 1.1 Flash is live in Picsart, and its headline number is 4K. The first Gemini Omni topped out at 720p, and this version adds 1080p and 4K, so a clip can leave the model at a size that satisfies a real delivery spec.<\/p>\n
Resolution is the news, but it is not the reason Omni is unusual. Most AI video models return a silent clip that somebody else has to score, caption, and cut. Omni was built to hand back something closer to finished, and the sections below cover the three behaviors that make it feel that way.<\/p>\n
Audio is generated alongside the frames rather than layered on once the render completes. There is no separate voice engine to line up and no effects library to borrow from.<\/p>\n
That shared origin is what makes the output cohere. Footsteps land on the frame where the foot lands. Spoken lines match the mouth shapes producing them. Room tone belongs to the space on screen and stays there as the shot moves. Speech, music, and atmosphere all trace back to the same instruction that described the picture, and it is why the model takes no audio file at all: whatever should be heard belongs in the written brief.<\/p>\n
A finished clip can be changed by describing the change. No timeline, no masking, no starting over.<\/p>\n
The instructions really are that ordinary. Swap a red car for a black one. Take a logo off a shirt. Recolor a jacket. Make a delivered line sound more apologetic. Omni rewrites the frames that the request touches and leaves the rest pixel-stable, so whatever was already correct stays correct.<\/p>\n
That turns a round of notes into a round of sentences, which is a different working rhythm from re-rendering the whole clip and hoping the fix lands this time.<\/p>\n
Type is where AI video usually gives itself away, with letters that warp between frames or spell a word two ways in four seconds. Omni holds typography steady instead.<\/p>\n
Captions keep their shape. A call to action on an ad stays legible at the end of the clip as well as the start. An equation on a blackboard, a label in a product demo, a lower third on an explainer: all of it survives the motion. Video that has to carry words stops requiring a separate design pass to fix them.<\/p>\n
Resolution is not the whole update. A single generation now accepts up to five reference images and three reference videos, sets its own start and end keyframes independently, and continues footage you already have instead of starting over. Clips run at any whole second between three and ten, in landscape or portrait.<\/p>\n
The first Gemini Omni took one reference image and offered no keyframe slots, so a camera move had to be described rather than shown. This is the version you can direct.<\/p>\n
The fastest way to judge a one-pass model is to close your eyes for the second half. Generate this, then listen to whether the sound belongs to the picture or merely accompanies it.<\/p>\n A blacksmith brings a hammer down on glowing metal in a dim workshop, sparks scattering across the anvil, bellows breathing in the background. Each hammer strike rings out and decays into the stone room, the fire hisses when the metal is quenched, and a low murmur of wind passes the open door between blows<\/p>\n\n One-pass audio test<\/h3>\n\n