Generate AI videos from text prompts using Google Veo

This promptfoo example demonstrates Google's Veo video generation models for AI-powered video creation from text prompts, source images, or a combination of both together. It covers five models - veo-3.1-generate-preview (latest, with video extension support), veo-3.1-fast-preview, veo-3-generate, veo-3-fast, and veo-2-generate - each supporting 4, 6, or 8 second clips (veo-2 also supports 5 seconds).

Use this example when you want to evaluate Veo's text-to-video, image-to-video, or video-extension capabilities. It works through either a Google AI Studio / Gemini API key or a Google Cloud project with Vertex AI enabled and gcloud application-default login - video extension specifically requires the explicit Vertex provider path and an operation ID from a prior generation, and is only available on Veo 3.1.

Configuration options control aspect ratio (16:9 default or 9:16), resolution (720p default or 1080p), durationSeconds, personGeneration (allow_adult or dont_allow), a negativePrompt to steer away from unwanted concepts, an image input for image-to-video, a lastImage for interpolation to an end frame, extendVideoId for extending a prior Veo 3.1 generation, and up to three referenceImages for style reference:

npx promptfoo@latest init --example google-video
cd google-video
npx promptfoo@latest eval

Text-to-video uses promptfooconfig.yaml, image-to-video uses promptfooconfig-image.yaml, and video extension uses promptfooconfig-extension.yaml.

Set GOOGLE_API_KEY for AI Studio, or enable the Vertex AI API and set GOOGLE_PROJECT_ID for Vertex AI. google:video:* uses AI Studio by default and auto-detects Vertex AI when project-based auth is configured; existing project-based google:video:* configs remain compatible, but vertex:video:* is the recommended explicit path for Vertex-only features like extendVideoId. Generated videos are stored in promptfoo's blob storage with content-addressable hashing for automatic deduplication, served via the local server for viewing, and can be forced to regenerate with --no-cache; Veo generation itself runs as a long-running operation with polling for completion.

Teams evaluating AI video generation quality and features - text-to-video, image-to-video, style reference, or extending existing clips - across Google's Veo model lineup. Veo 2 clips run 5, 6, or 8 seconds while the Veo 3.x family runs 4, 6, or 8 seconds, so teams comparing durations across generations should account for that difference rather than assuming identical duration options everywhere. The lastImage option (an end frame for interpolation) and the up-to-three referenceImages for style guidance are worth testing separately from plain image-to-video, since each shapes the output differently.