Clips & Sources

The Clipsstep of a project lists its sources. A source is either a video you upload or a generated film you add from the project's exports. Once prepared, a source never changes: every clip cut from it pins the exact file (its fingerprint) and version, so a re-export or a new upload becomes a new source rather than silently changing old clips.

• Rights first. You must confirm you own the video or hold the rights to clip and publish from it before an upload can start. The acknowledgement is recorded on the source.
• Declared limits. The accepted types, the maximum size and the maximum length are enforced by the server and shown on the page; a file that breaks them is refused before it uploads (or, for length, as soon as it is probed).
• Resumable uploads. Large files go up in chunks. If the connection drops or the page reloads, choosing the same file again resumes from the chunks the server already holds; Pause keeps the session for later. Nothing is assembled until every chunk has arrived.
• Preparation. After the upload (or when a film is added), Filmgenic probes the file, makes a browser proxy and a poster, and transcribes speech with word timings when a transcription provider is configured. A film with no speech is still a valid source; the card says what happened.
• No links. Sources come from your files and your project — there is no "paste a URL" import, by design.
• Retention. Sources are kept for the retention period shown; delete one from the card when you no longer need it.

Open Clipson a prepared source, pick a target length and choose "Find clips". Filmgenic proposes ranges from the source's own structure and says why each one was chosen:

• Complete thoughts. When the source has a transcript, a candidate starts on a sentence start and ends on a sentence end, close to your target length, preferring clean pauses on both sides. It never begins or ends in the middle of a word.
• Whole shots. A generated film without speech is cut on its own edit points: runs of whole shots between cuts.
• Evenly spaced windows. With neither, Filmgenic says so and offers evenly spaced windows to adjust by ear rather than pretending to judge.
• Your edits win. Play the range on the proxy, move the start or end (snapped outward to word boundaries when there is a transcript), keep or dismiss. Kept and dismissed clips survive when you find clips again with another length; untouched proposals are replaced.

The order is a heuristic about length and boundaries for this source only. It is not a prediction of how a clip will perform anywhere.

Framing & captions on a candidate opens the clip editor. Pick the shape you are cutting for (9:16, 1:1, 16:9) and the green box shows exactly what that clip will contain; the dashed box inside it is the caption-safe area, roomier at the bottom on a vertical clip where players put their own chrome.

• The crop follows the subject. Play the clip, click the picture where the person is, and that becomes a keyframe. Between keyframes the crop moves smoothly, before the first and after the last it holds still, and it is always clamped inside the frame, so a subject who walks across the shot stays in the picture.
• A film with no dialogue is cut on its shots. Before anything is transcribed, Filmgenic listens to a few moments of the audio and decides what is actually there: speech, singing, or neither. A score-only film is never sent to a speech recogniser — handed music, a recogniser does not return silence, it returns confident invented sentences — so instead the source is scanned for its cuts and clips start and end on whole shots. The source says which happened, and captions say plainly that there is nothing to caption.
• Or let it find them. "Track the subject" samples the clip, works out where the person or product is in each sampled frame, and writes the keyframes for you. It is a starting point, not an authority: the result is never marked hand-corrected, so any click of yours still wins, and it refuses to overwrite framing you set by hand unless you ask it to. A detection that jumps impossibly far between samples is reined in rather than followed, so the crop glides instead of snapping.
• Your corrections stick. A track you have saved is marked hand-corrected and is never rewritten automatically. Zoom in for a tighter frame; the crop size stays constant across the clip so the movement never wobbles.
• Two speakers. Tick "Two speakers, stacked" and the vertical clip becomes two halves, each following its own person, instead of one wide shot with two small heads.
• Captions from the words. "From transcript" writes the lines for the words actually inside the range, breaking on sentence ends and real pauses rather than at a fixed length. Edit any line; an edited line is kept when you move the clip's start or end, and the untouched ones are re-cut to match.
• Poster frame. Pause where the clip looks best and choose "Poster frame here".

Renders turns the clip you edited into an actual file: vertical 9:16 (1080×1920), square 1:1 (1080×1080) or wide 16:9 (1920×1080). The crop follows your keyframes, the captions are burned in, and the poster frame becomes the thumbnail.

• A render belongs to one edit. Asking twice for the same clip, framing and captions gives you the same file. Change the range, the framing or a caption and the next render is a new file: a link you already shared keeps playing what it played before.
• It runs in the background. The row shows queued, rendering, ready or failed, with the size and length once it is done, and the reason if it failed. Asking again for a failed render starts it over.
• Nothing is published. A render is a file you can download. Sharing and publishing are a separate step.