Gemini Omini 1.2 Tops the Charts with $0.07-per-Second 360p Video and 4K Upscaling

Recently, Google DeepMind has released a new generation of video generation models:Gemini Omni 1.2 Flash。

According to Arena.ai, Gemini Omni 1.2 Flash is ranked 1st in both text-generated videos and image-generated videos and 3rd in video editing.

Gemini Omni 1.2 Flash is ranked No. 1 in text-generating videos with 1,541 points, up 27 points from the previous Gemini Omni 1.1 Flash, which was ranked No. 2 with 1,514 points.

With this release, Google's Omni series of models encompasses the top three producers of text-generated videos.

Google's update focuses on:

The duration of the video has been increased to 15 seconds, and source clips of up to 15 seconds can be edited.

Video length from 3 seconds to 15 seconds, supporting native audio

Source videos of 3 to 15 seconds can be edited, supporting up to 7 reference images

Each call can be extended by 3 to 10 seconds, with a maximum of 30 seconds.

Any output content can be re-edited in subsequent calls via videoId

It supports 360p, 720p, 1080p and 4K, with a 16:9 or 9:16 aspect ratio.

This time, Google doesn't play "Queen Application" and directly uploads the full Google AI Studio and Gemini API.

Developers, corporate customers and ordinary creators can call directly from behind the scenes.

This update, which was 15 seconds longer than its predecessor, saw a 30% drop in sales compared to the previous generation. In addition, Google added new features such as multi-round dialogue tuning, film-level stability cameras, and industry-level "draft-refining" workflows.

In the "Blind Box Era" of the AI video, Google personally pressed the end button.

Over the past year, the field of AI video generation has seemed extraordinary, but anyone who actually uses it has a bitter feeling:

The cost of drawing cards is very high, modification is the equivalent of drawing again, images cannot be mass-produced.。

Almost every update of the Gemini Omni 1.2 Flash released by Google is breaking down these pain points.

In the past, the workflow was: write a good cue word, directly hardening the scalp and rendering 1080p. It was not until it was generated that the character's fingers were one more, or the protagonist's vision was distorted.

This whole calculation, even if it is completely discontinued, you have to pay the full amount.

The new Gemini model directly overturns this billing logic.

Its core solution, called "low-cost modeling, on-demand upgrades":

In addition to being cheaper, this workflow can help you control costs: you can quickly generate an original in 360p or 720p, then edit and modify the details without having to re-generate the entire video, and finally upgrade the most satisfying work to 1080p or 4K.

This logic completely changed the cost of trial and error for the team.

In the past, a 15-second commercial demo, 20 repeated attempts and mistakes, could cost hundreds of dollars.

Now, all early proof of concept has been squeezed at a draft level of $0.07 per second, with the total cost squeezed down to a third or even less than the original.

For advertising creative companies, offshore e-commerce vendors and individual video creators, this means that for the first time, AI video has the economic feasibility of a "large-scale industrialization trial and error".

The new Gemini model brings new features:Multi-turn Conversational Editing is supported by Native。

This means that creators don't have to leave the dialog box to quickly derive multiple ad variants for different social media audiences in a few rounds of interaction.

Change doesn't have to be pushed back, productivity is the real liberation.

Gemini's new model is getting stronger.Reference Adherence。

Now, you can enter multiple reference points into the model at the same time in addition to the prompts:

portraits of faces or three-dimensional views of characters;

the white background of the product or the design of the packaging;

The longest.3-second movement reference clip；

Audio clues or rhythm points.

According to the official announcement, the new model is specifically reinforcedControl of depth film footage：

Whether the drones pass through a narrow gate into the old building or slide through the ground at high speed, the footage is flat, through a sharp mesh of silk.

In terms of understanding three-dimensional spatial relationships, there is also a qualitative leap, movement processes coherent with nature, and lens movement with a more professional film industry texture.

In Gemini's new model, the multi-character dialogue scene has led to significant improvements:

The pronunciation of the mouth is closely aligned with the tone, and the lip muscles are more realistic.

This is the update.The system supports multiple characters speaking in sequence in the same image, allowing the system to accurately distinguish the speaker's subjects without the confusion of "character A opening the mouth, character B issuing the voice".

Whether it's in conjunction with Google Vids to train corporate employees in short videos, explaining explanatory videos, or creating multi-character dialogue animations, the combination of sound and images is closer to real photography.

Within hours of the release of the new model, a series of demo tests have been produced by developers and visual creators around the world, and have been hotly discussed on social media platforms.

The whole web test, in which several groups of representational contrasts are very clear, illustrates the problem:

1. Battlefield FPV Operating Mirror vs. Omni 1.2 vs. 1.1)

DesignArena compares the two recent versions with the same pronoun, with the lens rapidly plunging from the outside into the inner circle of the Roman battlefield.

The previous version 1.1 showed clear image blinking and pivotal distortion as it passed through the shadows of the vertical column; while 1.2 Flash's entire traversal process was smooth, with no rupture at the intersection of the aperture and the optical shadow, and the stability completely exceeded that of its predecessor.

2. Testing of single-word wildlife documentaries

A single natural-language prompter can be used to produce almost randomly real wildlife documentaries.

The netizens shouted:

The reality, the animal movements and the sound effects are incredible.

3. Pressure test for action and role consistency

Blogger Masamune_sakaki uses the Gemini Omni 1.2 Flash to try and get Delta to dance:

His assessment was: "Jump up, and the character design can stay".

He also added that if the cost is appropriate, you can consider making a full-length animation.

He also tried Deltamon's Running Cool video, and not only did the design not break down, but the background was great!

Each character has a completely different artistic style, and Gemini Omni 1.2 Flash has done exceptionally well:

Kosla is the animation genre, Kimong is the 3D writing genre, Spider-Man is the manga genre, Picchu is the animation genre, and there are several other characters that all interact in the same scene.

Here's a look at the evolution of AI video over the past few years:

In the first phase, we're working together."I can't move".The people in the picture can turn their heads down, blink their eyes, and create a whole network of surrounds.

In the second phase, we're working together.It's a surprise of quality.Anyone who can generate a high-resolution camera in a few seconds with a hint word will be able to browse the tech circle.

Today, the launch of the new Gemini model officially announced that the AI video has been pushed to phase three:The price of rolls, roll delivery, roll workflow adapted.