Skip to content
GenMedia Creative Studio

Veo Video Generation

The Veo workspace (/veo) is GenMedia Creative Studio’s flagship video synthesis environment. It exposes all capabilities of the Vertex AI Veo model family (veo-2.0, veo-3.0, and veo-3.1) through a capability-driven UI backed by asynchronous Long-Running Operation (LRO) polling and an async FastAPI router.

Veo Interaction Architecture

Source Architecture

Section titled “Source Architecture”
Layer File Path Responsibility
Mesop Page & State pages/veo.py, state/veo_state.py (PageState) Multi-mode UI workspace, reference image/video uploaders, camera & cinematography controls, and video player gallery.
UI Components components/veo/generation_controls.py, components/veo/file_uploader.py, components/veo/video_display.py Modular mode switchers, dynamic parameter selectors bound to config/veo_models.py, and C2PA provenance viewer.
Model Adapter models/veo.py (generate_video, VideoGenerationRequest) Dispatches client.models.generate_videos(), polls LRO completion, stores local byte responses to GCS when needed, and extracts C2PA manifests.
Async FastAPI Router routers/veo_router.py (/api/veo/generate_async, /api/veo/TaskStatus/{task_id}) Non-blocking background task execution for long-running video generations.
Model Registry config/veo_models.py (VEO_MODELS, VeoModelConfig) Declarative capability registry defining supported modes, aspect ratios, durations, resolutions, audio support, and per-mode overrides for every Veo model.

Supported Generation Modes

Section titled “Supported Generation Modes”

The workspace adapts its upload slots and parameter constraints dynamically based on the active mode (supported_modes in config/veo_models.py):

  1. Text-to-Video (t2v): Synthesizes video directly from a descriptive text prompt, negative prompt, and camera direction.
  2. Image-to-Video (i2v): Animates a single starting reference frame (reference_image_gcs) into a full video clip while preserving the subject and composition.
  3. First & Last Frame Interpolation (interpolation): Accepts both a starting frame (reference_image_gcs) and an ending target frame (last_reference_image_gcs), synthesizing a smooth cinematic transition between the two keyframes (GenerateVideosConfig.last_frame).
  4. Reference-to-Video (r2v): Accepts up to 3 character/subject ASSET reference images (r2v_reference_images) and/or a STYLE reference image (r2v_style_image) via VideoGenerationReferenceImage, maintaining character identity or visual aesthetic across newly generated scenes.
  5. Video Extension (extend): Extends an existing Veo video (video_input_gcs) by an additional 7s segment using veo-3.1-generate-001 or veo-3.1-fast-generate-001.

Model Matrix (config/veo_models.py)

Section titled “Model Matrix (config/veo_models.py)”
Version ID Model Name Supported Modes Resolutions Durations Native Audio
3.1 (Default) veo-3.1-generate-001 t2v, i2v, interpolation, r2v (ASSET), extend 720p, 1080p, 4k 4s, 6s, 8s (7s on extend) Yes
3.1-fast veo-3.1-fast-generate-001 t2v, i2v, interpolation, r2v (ASSET), extend 720p, 1080p, 4k 4s, 6s, 8s (7s on extend) Yes
3.1-lite veo-3.1-lite-generate-001 t2v, i2v, interpolation 720p, 1080p 4s, 6s, 8s Yes
3.0 / 3.0-fast veo-3.0-generate-001, veo-3.0-fast-generate-001 t2v, i2v, interpolation 720p, 1080p 4s, 6s, 8s Yes
2.0 / 2.0-exp veo-2.0-generate-001, veo-2.0-generate-exp t2v, i2v, interpolation, r2v (ASSET & STYLE on 2.0-exp) 720p 5s–8s No

Generation & LRO Polling Lifecycle

Section titled “Generation & LRO Polling Lifecycle”
  1. Request Validation (VideoGenerationRequest): Validates prompt presence, resolution (720p/1080p/4k), aspect ratio (16:9 or 9:16), video count (1–4), person_generation policy, deterministic seed, and optional enhance_prompt / generate_audio flags against the selected VeoModelConfig.
  2. Vertex AI LRO Submission: models/veo.py constructs a types.GenerateVideosConfig with output_gcs_uri=cfg.VIDEO_BUCKET and invokes client.models.generate_videos(model=..., prompt=..., image=..., video=..., config=...).
  3. Polling Loop: The adapter polls client.operations.get(operation) every 10 seconds until operation.done is True.
  4. Post-Processing & Provenance:
    • If the API returns inline video_bytes instead of a gs:// URI, models/veo.py uploads the MP4 bytes directly to cfg.VIDEO_BUCKET.
    • C2PA Content Credentials: When present in the generated MP4 container, C2PA manifest data is extracted via common/c2pa_utils.py and attached to the Firestore MediaItem (c2pa_manifest) for cryptographic provenance inspection in the UI.