Your own keys take priority; if your account has server keys enabled, missing ones fall back to the server defaults. Keys are stored on the server and never shared.
Saved
Server keys are enabled for your account. Missing keys fall back to server defaults.
API Usage & Costs
Estimated using published list rates. Cross-check against provider dashboards for exact billing.
Pick an AI avatar to host the video on camera, intercut with b-roll — or keep "No presenter" for a faceless video.
Pair "Bookend presenter" with the "Bookend / Gary style" quality preset in Step 3 (calm presenter, slow-motion b-roll, Budget models).
55%
Scenes are cut at scene level (segmentation is set to scene-driven) so the presenter and b-roll can alternate every few seconds.
Segmentation
Parsing Progress
Parsed Scenes
Voice Settings
0.9
Timing
0s
0s
0s
0s
Manage My Voices
Progress
Voice Previews
AI Avatar Presenter
Needs a Gemini key. yt-dlp searches YouTube, Gemini picks the moments that show each scene's subject, only those seconds are downloaded at 1080p, and Gemini re-checks every cut. Sources are credited in the generated description.
Prompt-driven: OmniHuman and InfiniteTalk follow it best, Kling Avatar partially, VEED Fabric ignores it. Regenerate a clip with a motion note for one-off fixes.
Each on-camera scene becomes a lip-synced clip driven by that scene's voiceover. Viewers decide in the first minutes, so the premium model goes there. The settings below only apply to b-roll scenes.
Visual Settings
Applies model choices below; edit anything afterwards and the preset switches to Custom. The estimate under the button shows the per-minute cost.
Applies to presenter clips, animated b-roll and real footage alike. Most YouTube footage and most video models run at 24 or 25 fps; duplicating frames onto the 30 fps timeline is what makes assembled video judder.
Real footage is selected: clips come from YouTube and the public-domain FedFlix archive on archive.org (the library behind destockd.com), not from an image model. The image settings are hidden; the model chosen there () is only used for picture-in-picture insets and as a fallback still when no usable footage is found. Show image settings
Manage My Styles
60%
Reference channels cut to motion clips for roughly 60% of b-roll and hold Ken Burns stills for the rest; action scenes get animated first.
Progress
Visual Previews
Render Options
Image Motion (Ken Burns)
Only affects how the talking-head clips are framed in the edit, so changing it just needs another render.
Gemini watches the whole video and flags frozen frames, visuals that don't match the narration, audio glitches, caption problems and rendering artifacts.