VideoGenerationModel generates videos with Google's Veo models on Vertex AI. Generation is a long-running operation: you start it, then poll until the video is ready.

Veo on the Gemini API is retired. On the Gemini API, VideoGenerationModel.GenerateVideosAsync throws NotSupportedException straight away instead of sending a request the API would reject. To generate video on the Gemini API, use a Gemini Omni model (GoogleAIModels.Video.GeminiOmni12Flash, or InteractionModels.GeminiOmni12Flash in the Interactions API) through the Interactions API.

NuGet packages

dotnet add package Google_GenerativeAI
dotnet add package Google_GenerativeAI.Auth   # service account / OAuth authenticators for Vertex AI

Video generation is billed per second of output and can be expensive; start with short durations and one video.

The GoogleAIModels.Video.Veo* constants are marked [Obsolete] because those models are shut down on the Gemini API. Check the Vertex AI documentation for which Veo versions are currently available in your region.

using GenerativeAI;
using GenerativeAI.Authenticators;
using GenerativeAI.Types;

var vertexAi = new VertexAI(
    projectId: "your-project-id",
    region: "us-central1",
    authenticator: new GoogleServiceAccountAuthenticator("service-account.json"));

var videoModel = vertexAi.CreateVideoGenerationModel(VertexAIModels.Video.Veo31FastGenerate001);

var operation = await videoModel.GenerateVideosAsync(
    "A dog catching a ball in a sunny park",
    config: new GenerateVideosConfig
    {
        AspectRatio = VideoAspectRatio.LANDSCAPE_16_9,
        DurationSeconds = 5,
        NumberOfVideos = 1,
        Resolution = VideoResolution.HD_720P,
        PersonGeneration = VideoPersonGeneration.AllowAdult,
        EnhancePrompt = true
    });

var result = await videoModel.AwaitForLongRunningOperation(operation!.Name!);

if (result?.Done == true && result.Result?.GeneratedVideos is { Count: > 0 } videos)
{
    var video = videos[0];
    if (video.VideoBytes != null)
        await File.WriteAllBytesAsync("dog.mp4", video.VideoBytes);
    else
        Console.WriteLine($"Video stored at {video.Uri}"); // when OutputGcsUri is set
}

You can also construct the model directly: new VideoGenerationModel(platformAdapter, VertexAIModels.Video.Veo31FastGenerate001), where platformAdapter must be a VertextPlatformAdapter.

GenerateVideosOperation has Name, Done, Error and, when done, Result (GenerateVideosResponse):

• GeneratedVideos: a List<Video>; each Video has VideoBytes (also exposed as BytesBase64Encoded), Uri and MimeType.
• RaiMediaFilteredCount and RaiMediaFilteredReasons: videos removed by safety filters.

Pass a starting frame as an ImageSample (bytes or a Cloud Storage URI):

var firstFrame = new ImageSample
{
    ImageBytes = await File.ReadAllBytesAsync("first-frame.png"),
    MimeType = "image/png"
    // or: GcsUri = "gs://bucket/first-frame.png"
};

var operation = await videoModel.GenerateVideosAsync(
    "The camera slowly pulls back to reveal the whole city",
    referenceImage: firstFrame,
    config: new GenerateVideosConfig { DurationSeconds = 5 });

These settings are sent to Vertex AI by VideoGenerationModel:

GenerateVideosConfig also defines GenerateAudio, LastFrame, ReferenceImages (VideoGenerationReferenceImage with VideoGenerationReferenceType.STYLE/ASSET), Mask (VideoGenerationMask with VideoGenerationMaskMode) and CompressionQuality (VideoCompressionQuality.OPTIMIZED/LOSSLESS). VideoGenerationModel does not currently forward these to the Vertex AI request.

• Image Generation
• Interactions API (Gemini Omni video on the Gemini API)
• Authentication
• Supported Models