Veo Omni Flash 1.2 Rises from 7 to 8.5 in One Test, but Mixes Up References

Video models are usually praised based on demo clips the vendor selected itself. That's why it's more interesting when someone takes a new version, runs it on the same prompts as the previous one, and honestly reports where it stumbles. This is exactly how a Habr author tested Google Veo Omni Flash 1.2.

What the tests showed

The author has two rankings of his own. In the first, "text-to-video," 12 models received the same prompt about a street cook with noodles, and he rated each from 1 to 10. Veo Omni Flash 1.2 scored 8.5 and took 4th place. The previous version, 1.1, got 7 points and landed in 6th place. In both cases a clip costs 21 tokens, meaning quality improved without a price increase.

In the second test, "image animation," the model receives three ready-made images (a portrait, a construction site, and a jar) and has to bring them to life. The result is scored out of 100. Omni Flash 1.2 got 76.2, and version 1.1 got roughly the same. So if your main use case is animating a finished picture, the upgrade gives you almost nothing.

Both tests were repeated on the same prompts and frames, so the comparison between versions is fair. This is a strong point of the review: only the model changes, not the conditions.

New: 15 seconds and 7 references

Clip length has grown to 15 seconds. In addition, in SYNTX.AI you can upload up to 7 images in reference mode. References are needed so that characters and objects look the same from scene to scene. For advertising and serial formats, this matters more than a pretty picture.

In practice the results were uneven. In an ad with a lineup of 6 packages, the model held on to 4. In a clip about a gym with 7 trainers, details slipped for three of them: hair, a tattoo, position in the frame. In other words, the claimed "7 references" does not mean all seven will come out recognizable. The more objects in the frame, the higher the chance that some of them will "drift."

Physics: the weak spot

The author separately notes the physics. On the jar, droplets stand still even though they should run down. And a punching bag swings even though nobody is hitting it. The first error is a lack of motion, the second is motion without a cause. Both are noticeable even without a frame-by-frame breakdown, and both are easy to miss if you watch the clip once on a small screen.

What to do with this

• Text-based scene drafts. For the same 21 tokens, the new version is rated higher than the old one. If you are storyboarding an idea or testing plots, it's worth switching.
• Ads with liquids. Droplets, spilled drinks, steam: check them frame by frame. According to the author's observations, the model errs here more noticeably than Google promises.
• Product lineups and groups of people. Don't count on the model holding all the references. Split the scene into several clips with fewer objects and stitch them together.
• Animating ready-made frames. If you're already on version 1.1 and mostly work with "image animation," there's no urgent reason to move.

Limitations of the review

The scores are subjective: they were given by one person. The text test is built on a single prompt about a cook with noodles, so 8.5 versus 7 speaks to a specific scene, not to all genres. The price of 21 tokens refers to the SYNTX.AI platform, not to Google's direct API. The author added the reference scenes only for version 1.2, so there is no direct comparison with 1.1 on them.

The bottom line is simple: Veo Omni Flash 1.2 has become better for quick text-based drafts, but it has not become reliable for ads with liquids and for frames with lots of characters. If you make production clips, budget time for checking and regeneration, and don't trust the "up to 7 references" figure.