What we're building
PiGenTek AI Studio makes video creation accessible — no filming, no editing software, no production team. But "accessible" and "easy to say" are different from "actually hard to build." Underneath that simple promise are a set of genuinely unsolved problems in AI video generation: keeping a character's face consistent across multiple shots, making a story feel like one continuous scene instead of stitched clips, giving a generated video a voice that sounds like it belongs to someone.
Creating longer cinematic sequences from simple prompts
The video models we build on can only generate a few seconds of footage in a single call — that's a real, current limit of the technology, not a choice we made. So a 30-second or 90-second clip isn't one generation — it's several shorter scenes, planned to continue from each other and joined together afterward.
The harder part isn't generating the individual scenes. It's making them feel continuous — right camera variety so a long clip doesn't feel like the same shot repeated, right pacing so scenes don't feel disconnected, and doing all of this without the process becoming so heavy that it's unreliable. We've rebuilt this pipeline more than once this year to get the reliability right — including moving where the final assembly happens, after learning the hard way that piling too much heavy processing into one long server request is a fragile way to build this.
Motion and scene continuity
A long clip made of several separately-generated scenes can easily look like exactly that — disconnected shots with no relationship to each other. Two things help: planning deliberate shot variety scene-to-scene (a wide shot, then a close-up, then movement — not the same framing five times in a row), and carrying lightweight context between scenes so each one knows roughly what happened right before and right after it.
Character consistency
Ask an AI video model to generate the same person across five separate scenes, and by default, you won't get the same person — the face, the outfit, sometimes even the apparent age can drift shot to shot. We anchor character identity using dedicated reference portraits — one isolated image generated for each recurring character, then reused across every scene that character appears in, including when several recurring characters share the same scene together.
This follows the same "ingredients" pattern the underlying model itself documents for keeping people, objects, and settings visually consistent across a generation — we're not working around a limitation, we're using the tool the way it was actually designed to be used.
Lip-sync and dialogue
This is the one area on this page we're being deliberately cautious about. We evaluated an approach to precise, multi-character lip-sync and weren't satisfied with the result — the mouth movement didn't hold up to the standard we want for something we'd put in front of clients. Rather than ship something that looks close-but-not-quite-right, we're treating this as an open problem, not a solved one. Single-speaker narration, where a video simply needs to sound right rather than have every word visually mouth-matched, is a different and more solved problem — that part works today.
AI narration and voice
Every clip needs a voice, and we support two paths: a set of professional narration voices for straightforward use, and a "My Voice" option where you can record or upload a short sample, and we blend it with your chosen style and delivery to approximate your voice's character. Worth being precise about what this actually is: it's not literal voice cloning — the underlying model doesn't offer that — it's a best-effort simulation using your sample as a reference alongside a detailed style prompt. Genuinely useful, honestly described.
Multi-scene storytelling
Longer stories need more than continuous motion — they need narration that reads as one throughline instead of several disconnected lines, and a planning step that thinks about the story, not just a list of shots. We're building toward this by having the same planning step that breaks a story into scenes also handle character locking and narration pacing together, rather than treating them as separate problems bolted together after the fact.
What we're testing next
Curious what this looks like in practice? Want to put the technology to work?
Explore PiGenTek AI Studio →