Hi Google Cloud Vertex AI & Gemini Omni Team,
We are using the new Gemini Omni 1.1 Flash Video model via the official Interactions API to generate sequential multi-scene videos using the Scene Extension feature (previous_interaction_id).
We encountered a reproducible issue where interaction status remains stuck in ‘in_progress’ for an extended period (> 60 minutes) on 30s cumulative video, which in turn causes immediate HTTP 500 errors when attempting to chain the next scene.
Here are the technical details and reproducible interaction IDs from our GCP project:
- GCP Project ID: (PII Removed by Staff) (Project Number: [PII Removed by Staff])
- Region: global
- Model: gemini-omni-1.1-flash-preview
- Endpoint: https://aiplatform.googleapis.com/v1beta1/projects/project-PIIRemovedbyStaff/locations/global/interactions
Observed Timeline & Interaction IDs:
-
Scene 1 (10s):
- ID: ChAxYWRiNDFjZDY4YWEwMjUyEAgaATAqBG1haW4
- Status: ‘completed’ (transitioned almost immediately).
-
Scene 2 (20s cumulative, chained from Scene 1 via previous_interaction_id):
- ID: ChA0M2QxMzdkMzE4ZmRmODZkEAgaATAqBG1haW4
- Status: ‘completed’ (took ~15 minutes to transition from ‘in_progress’ to ‘completed’).
-
Scene 3 (30s cumulative, chained from Scene 2 via previous_interaction_id):
- Run 1 ID: ChBjMTllYmVhODA2NGFhYWIzEAgaATAqBG1haW4 (created at 11:49 UTC)
- Run 2 ID: ChA4YmUyMjg5NzBiM2VmZDk1EAgaATAqBG1haW4 (created at 12:08 UTC)
- Observation: In both runs, the 30s cumulative video was successfully generated and returned in the client response. However, when querying the interaction status via
client.interactions.get(), the status remains stuck in ‘in_progress’ for over 60 minutes.
The Error:
When attempting to generate Scene 4 by calling:
client.interactions.create(model='gemini-omni-1.1-flash-preview', previous_interaction_id=SCENE_3_ID, ...)
The API immediately fails with:
500 api_error: Internal error encountered on Vertex AI Interactions service.
It appears the backend rejects extending from an interaction whose status has not yet transitioned to ‘completed’ in the metadata store.
Questions:
- What is the expected SLA/latency for a 30s or 40s cumulative video interaction to transition to ‘completed’? Why does metadata synchronization scale exponentially with video duration (10s: instant, 20s: 15m, 30s: >60m)?
- If the video output data has already been delivered to the client, could the Interactions API allow
previous_interaction_idchaining without blocking on background KV-cache flush, or at least return a clear dependency error code rather than an unhandled HTTP 500?
Thank you for your assistance!