Repeated image-to-video audio failures on Veo 3.1 Lite and Fast

Gemini API
1

Hello Google Gemini API team,

We need help diagnosing three completed image-to-video requests through the Gemini Developer API. All returned zero video samples and one filtered result. Attempts 002 and 003 returned an audio-issue explanation. Our attempt 004 recovery classified the returned explanation as PROVIDER_AUDIO_FAILURE; that label is our client’s classification, not a Google error code.

Operations:

  • 002: models/veo-3.1-lite-generate-preview/operations/dv23p8zu4iu3
  • 003: models/veo-3.1-lite-generate-preview/operations/yr6do061u1z9
  • 004: models/veo-3.1-fast-generate-preview/operations/0s6gzjfqm7lt

Settings: durationSeconds=4, resolution=720p, aspectRatio=9:16, sampleCount=1, personGeneration=allow_adult.

Attempt 004 endpoint:
https://generativelanguage.googleapis.com/v1beta/models/veo-3.1-fast-generate-preview:predictLongRunning

Request shape: one instances entry containing prompt and image; image uses bytesBase64Encoded and mimeType=image/png. The base64 decodes successfully. The image is an adult fictional male seated in an office.

Prompt used for 003 and 004:
“Animate this adult fictional character seated in the office. Preserve his appearance, clothing, lighting and background. Use a fixed camera with subtle breathing, one natural blink and a slight head turn. Four seconds, vertical composition. No dialogue or added text. Audio: soft, steady office room tone with a faint ventilation hum. No music.”

Expected: one playable four-second portrait video.
Observed: no video. Adding room ambience and switching from Lite to Fast did not resolve the audio failure. Each attempt was submitted once; automatic retries are disabled. The provider reported no charge; invoice reconciliation is unverified.

Please clarify:

  1. Do these operation IDs indicate an audio-processing problem, a specific content restriction, or another cause?
  2. Is the submitted image encoding and the 4-second image-to-video configuration supported on both endpoints?
  3. Is there a known issue affecting these operations?
  4. What supported corrective action should we take before another paid attempt?
  5. Can you confirm the billing outcome for these operations?

No API keys, account credentials, source image bytes, or signed authorizations are included in this report.