Every photo waits for the AI image check before anyone but its owner sees it, but a video would walk past it, since the model only looks at still images. Before a video is released, ffmpeg should pull its first frame, one frame every 20 seconds and one at every hard cut (scene detection on a 320-pixel copy, at most 24 frames) and send each through the existing scan; one rejected frame rejects the whole video and tells the member the second it happened at, and a scan that cannot run holds the video rather than releasing it. Members whose clip is refused get the notice photos already send; nothing else changes for them.
Where: Vutuv.Moderation.ImageScans, a new post_video_frame kind
An AI agent wrote this text in my name. I know that is problematic.
Every photo waits for the AI image check before anyone but its owner sees it, but a video would walk past it, since the model only looks at still images. Before a video is released, ffmpeg should pull its first frame, one frame every 20 seconds and one at every hard cut (scene detection on a 320-pixel copy, at most 24 frames) and send each through the existing scan; one rejected frame rejects the whole video and tells the member the second it happened at, and a scan that cannot run holds the video rather than releasing it. Members whose clip is refused get the notice photos already send; nothing else changes for them.
Where:
Vutuv.Moderation.ImageScans, a newpost_video_framekindAn AI agent wrote this text in my name. I know that is problematic.