Hi all — posting this here since it’s really two connected problems, both hitting a small solo project hard.
1. Repeated false-positive audio-safety rejections, no useful signal to fix them
Calling client.models.generate_videos() (Gemini Developer API, google-genai Python SDK, GOOGLE_API_KEY auth — not Vertex) against veo-3.1-generate-preview / veo-3.1-fast-generate-preview, one clip in an otherwise unremarkable prompt has failed three times in a row. The operation completes (done=True, no error) but returns no video, with this filter reason:
“We encountered an issue with the audio for your prompt, which means we could not create your video. This can sometimes happen due to our safety filters or other processing issues. Please modify your request and try again. You have not been charged for this attempt.”
The prompt is a 1940s historical-documentary animated scene — no violence, nothing graphic — of a man stepping through a gate and reacting to a street scene. I revised it three separate times, removing more content each attempt:
- Crowd described with some detail (“a loose crowd of travel-worn families… bundles and suitcases at their feet”) — failed.
- Reworded to distinct, individually-described people, no hardship language — failed identically.
- Removed the second party entirely, leaving just the main character reacting to an empty street — failed identically.
Camera movement, style suffix, aspect ratio, duration, and reference images stayed constant across all three. A different clip in the same episode, using the same character reference and a similar “reveal people waiting” beat but shown as background figures through a window rather than a face-to-face encounter, succeeded first try — so this may be about shot composition (something read as cueing dialogue?) rather than prompt content, but there’s no way to confirm since the filter reason gives zero actionable detail.
Also worth noting: generate_audio=False exists on GenerateVideosConfig but raises ValueError on the Developer API (“only supported in Gemini Enterprise Agent Platform mode”), and negative_prompt gets rejected with 400 INVALID_ARGUMENT: Negative prompt is not supported in your use case as soon as reference_images are attached. So there’s currently no supported way to steer or disable audio generation at all when using reference-image conditioning on the Developer API — even though the audio track gets thrown away in post anyway by anyone compositing their own narration.
2. False-positive rejections still consume the daily quota
This is the part that turns an annoyance into a real blocker: my project is Tier 1, capped at 10 requests/day for Veo 3 Generate. A safety-filtered rejection isn’t billed (per the pricing docs) — but it still counts as one of the 10 daily requests, same as a successful one. Three consecutive false positives on a single clip burned 30% of a whole day’s Veo access for zero output. Combined with 5 clips/day being a normal single-episode workload, Tier 1 is barely usable even without the filter issue, and effectively non-functional on a bad filter day.
Separately — still investigating on my end whether the daily quota window (reset apparently at midnight Pacific) may have attributed requests made near a day boundary to the “wrong” day, since usage felt similar to the prior day but the quota page showed a much higher peak. Flagging in case others have seen quota-day attribution look off.
Ask: any guidance on (a) reducing these false positives, (b) a way to disable/de-risk audio generation on the Developer API, or (c) whether filtered/no-video requests counting against RPD is intended behavior — would be really helpful. Happy to share exact prompts/reference images/request IDs if useful for repro.