From thought to timeline
A sentence — or a spoken idea — becomes a voiced, captioned first cut.
Turn an idea into a video
Say what it should be about, in your own words or out loud. The studio writes the script, finds the footage, adds the voice and captions, and hands back a first cut.
The first cut takes a few minutes. Check the rights on any footage before you publish.
Brief & output
Set the source, format and duration when you need more control. Start with the guided composer above for the fastest path.
Input
0 words · about 0 seconds of narration
One adaptive generation pass plans the brief, searches every compatible stock source in parallel, prepares image-motion fallback while discovery runs, and sends unresolved scenes to Wan 3.0. Narration remains the timing master. It is capped at sixteen scenes and still needs rights review before commercial use.
The long-form button queues the same strict path for a 3–10 minute fixture-only development export. It reports planning, selection, narration, assembly, and rendering stages while the browser polls; use a target duration from 180 to 600 seconds.
Automatic captions a portrait video a few words at a time — big capitals in step with the voice, the word being spoken lit only when its timing was measured — and keeps whole-sentence documentary subtitles for landscape and square video. Choose either to override the shape.
Every generated image or video is checked against its scene brief, and a video is checked on the file that will be shown, not only on its poster. The current visual acceptance floor is 0.51. If the reviewer is unavailable the job still finishes, but nothing is approved: the video is blocked and not rendered unless you ask for a degraded preview, which is labelled as one everywhere.
DeepSeek V4.1 Flash Fast verifies the selected pixels against each scene brief. Visual review is part of the enforced workflow and is never switched off by an outage; an answer it cannot give in a valid form is asked for again before it counts as unavailable.
The pipeline searches the configured sources, records rights evidence, and keeps review-required media clearly marked. It cannot be turned into an untracked legacy fallback from the browser.
The quality path fans out across every compatible configured source, ranks the combined per-scene shortlist, and uses stock-image motion or generated-image motion before Wan 3.0. Wan 3.0 is capped at 20% of total footage.
Recurring story guide
Optional for provider-quality story runs. Upload at least two local images, load the stored catalog, then enter their asset IDs here. The references must be rights-declared and are preserved by hash in the story manifest.
Load stored local assets to see eligible image IDs.
Account and usage
Checking whether the local account boundary is enabled...
This panel exercises the local account and entitlement boundary only. It does not charge a card or replace Stripe, an identity provider, TLS, or production payment webhooks.
Plan
Script editing
Revise the full narration or one planned scene before selecting media. Timing is recalculated and any assembled project is cleared because its assets may no longer match.
Load a real-media manifest to reuse existing narration and scene assignments, or use the neural-voice auto-select action directly with a stored catalog. The auto-select actions rank the supplied asset catalog against each scene's visual intent, reject weak matches, and can synthesize the selected configured TTS provider. Discover provider media for plan searches each scene's visual query and returns provenance-only candidates; acquire selected items explicitly before ranking or rendering. Missing or weak assignments fail explicitly.
Optional live stock search
Uses configured Pexels/Pixabay keys, plus explicitly enabled keyless Openverse/Wikimedia Commons/NASA adapters. Quality generation searches every compatible source in parallel and preserves each provider and rights status for DeepSeek verification. Results are provenance-only until rights review and local acquisition; they are not silently renderable.
Local uploads are hashed and stored with a rights sidecar. Video and audio are normalized before catalog insertion; review-required files remain unavailable to normal rendering until cleared.
Scenes
Human visual review
Inspect each selected shot against its narration. Submit a review before enabling strict visual rendering; this records an auditable attestation and does not clear media rights.
Recurring-character review
Inspect every character appearance against the approved guide, including identity, costume, props, and story state. Strict story rendering stays disabled until every appearance is reviewed.
Music
Choose a rights-declared audio bed, adjust its level, or mute it. The same setting is used by preview and export.
Brand and caption style
Choose an original caption treatment. The selected preset is stored in the project and used by preview and export.
Project JSON
Project storage
Optional local persistence. Start the API with --project-root to enable it.
Batch production
Submit up to five quoted jobs and inspect independent lifecycle state. This local panel exercises the batch coordinator; it does not claim that a coordinator transition has called a provider or rendered a video.
Queue prepared project batch becomes available after a real or fixture project is loaded and the server exposes the render queue. It uses the same manifest and quality-review evidence for each item.
Queue a prepared project
Advanced queue path for an already assembled project. The former direct render endpoint has been removed so every customer-facing generation uses the provider workflow or a bounded queue. Generating a fixture project fills this form automatically; you can also paste a manifest documented in docs/Real-media-pipeline.md.
Leave this off for local fixture QA. Turn it on for a release check; fixture-only or review-required assets will be rejected.
Premium generation is optional and separately metered. Load the fish quality demo first, choose a scene ID such as s2, generate one scene, then queue the returned manifest.