Google unveils Imagen 3 and Veo at I/O

Background

Text-to-image models take a written prompt and produce a picture. Two weaknesses have followed them for a long time. The first is artifacts: stray marks, distortions and smearing that pull attention away from what the picture is meant to show. The second is text: when a prompt asks for words inside an image, the model has to draw letters with the right shapes and the right spelling, not shapes that merely resemble letters. Video generation raises the difficulty again, because a model has to keep one scene coherent across many frames instead of getting a single picture right.

What happened

On 14 May 2024, at its Google I/O event, Google announced Imagen 3 and described it as its highest-quality text-to-image model at the time ^[1]. The same announcement introduced Veo, a generative video model able to create 1080p clips that run beyond a minute ^[1].

Key details

Google's claims for Imagen 3 centred on image quality and on text. The company said the model produced far fewer distracting visual artifacts than its prior models ^[1], and that it was its best model yet for rendering text, which it called a long-standing challenge for image generation models ^[1].

Veo was presented as producing high-quality 1080p video in a wide range of cinematic and visual styles ^[1]. Length matters here: holding a scene together for more than a minute of footage is a different problem from producing a few seconds of motion, because every frame has to stay consistent with the ones around it.

Neither model was opened to the general public at that point. Both went into private preview for select creators, Imagen 3 inside ImageFX and Veo inside VideoFX ^[1]. A private preview means a small group of users gets access before any broad release, so the models are exercised on real work while access is still controlled.

Google also said that videos generated with Veo inside VideoFX would be watermarked with SynthID ^[1]. A watermark of this kind leaves a marker inside the generated file, so that the output can later be recognised as machine-generated rather than mistaken for a camera recording.

The announcement was not limited to images and video. Google also showed Music AI Sandbox demos recorded with Wyclef Jean, Marc Rebillet and Justin Tranter ^[1], putting a third kind of generative media — audio — alongside the other two.

Why it matters

Google said Imagen 3 would come soon to Vertex AI ^[1], which points to a path from a preview for selected creators toward a model that developers can call on directly. That is the usual shape of a media-model release: a limited, hand-picked trial first, wider access afterwards.

The emphasis on text rendering is practical rather than cosmetic. Words inside a generated image only help if they are spelled correctly and legible; otherwise the image cannot be used for anything that carries a label, a title or a sign. Google presenting text rendering as a headline capability says that the models are being pitched at finished-looking material, not only at pictures whose overall look is convincing from a distance.

Watermarking belongs to the same story. Once a model can produce a minute of convincing 1080p video, telling generated footage apart from recorded footage stops being a technical curiosity and becomes a question about how the file itself is labelled. SynthID's role in the VideoFX output is to answer that question at the point of generation ^[1].

Finally, the grouping is itself notable. Image, video and music tools were introduced together, all framed around creators, which treats generative media as a family of related instruments rather than one isolated model launch ^[1].