Google: Veo 3.1 Lite

Veo 3.1 Lite is Google’s Lite member of the Veo 3 video-generation line. It synthesizes a video with audio from a natural-language description or an input image. The output is a generated clip, not a textual analysis of an existing recording. The maker’s Lite model card refers to Veo 3 for the underlying architecture, training material and known limitations, while the Lite-specific documentation defines its own generation features and limits.

The referenced architecture uses latent diffusion over temporal audio representations and spatio-temporal video representations. The base model was trained on audio, video and images, with captions at different levels of detail, and filtered and deduplicated training material. This explains the joined visual-and-audio generation purpose. It does not disclose a Lite-specific parameter count or establish that every generated movement follows physical reality. For a practical scene, decide the visible action and the corresponding sound together.

The exact Lite documentation gives a text-input limit of 1,024 tokens and one output video. The Lite column in the generation guide lists 4-, 6- and 8-second durations, with 720p as the default and 1080p requiring an 8-second clip. Landscape 16:9 is the default aspect ratio and portrait 9:16 is also listed; output is documented at 24 frames per second. Native audio is always on in that flow. These values belong to the producer’s named Lite workflow, not to an unrelated Veo edition or a separately measured application response.

Lite does not support 4K output or video extension. The parameter table also marks the multi-reference referenceImages feature and a video input for extension as unavailable in the Lite column. An initial image can establish the starting frame; the table lists a lastFrame image together with that initial image for a transition. Keep this frame-control path distinct from mixing several style or subject references. The producer’s people-generation settings are mode-specific: the Lite table lists allow_all for text-to-video and allow_adult for image-to-video and interpolation, while regional rules can further restrict that documented profile. Choose the starting image to match the first scene you actually want, with the subject, framing and visible details prepared before asking for motion.

For text-to-video, describe the subject, its action, the environment and the intended style. Add camera position and movement, framing, focus and light when they matter. An original product scene could keep the object centered while the camera approaches one material detail and the surrounding light changes gently. Identify the initial state, the main action and the final moment so the short clip resolves one coherent idea. A lengthy collection of unrelated actions is a different task from a focused shot.

For image-to-video, provide a clear initial frame and explain what should move and what should remain stable. Specify the subject’s motion and any camera movement separately. If using a final-frame constraint, prepare a compatible end composition and describe the transition between them. Inspect the result rather than assuming that a requested identity, position or shape remains identical throughout the clip. The referenced model card identifies complete consistency in complex scenes and complex motion as a continuing challenge.

The prompt guide supports explicit cues for dialogue, sound effects and ambience. Put exact speech in quotation marks, identify the speaker and describe sounds in relation to the visible action. A sound effect and background atmosphere should not be confused with a spoken line. English is documented as fully supported for prompting, while other languages have not been evaluated in that guide. Review the audible words, timing and sound quality as part of the same artifact as the video rather than treating audio as a guaranteed faithful transcript of the prompt.

The producer workflow is asynchronous: a generation request returns an operation, and the completed result is obtained after its status reports completion. Keep the operation identity and inspect the returned clip. The guide describes two-day producer-side retention and SynthID marking, so a retained artifact should include the downloaded result, prompt, source images and selected settings. A seed can improve repeatability but does not guarantee determinism. Check subject continuity, action, framing, sound and the intended endpoint of the scene before using the clip in an edit; processing completion alone is not proof that the creative requirements were met.