
A shot that's quick to generate can still be slow to finish. This article draws on two November 2026 production accounts, CJ 4DPLEX's ScreenX case study and BOT VFX's work on Young Washington, and proposes four measures that tell you more about a schedule than time to first image does.
Method. Sources reviewed November 18, 2026. This is a measurement proposal built from published interviews. We have not measured either production.
What this means for you
- Judge a tool on 20 shots from one sequence, not on your best 3. The worst shot sets your schedule.
- Stabilise every generated shot before you approve it. Warping that playback hides shows up once the shot is stabilised for tracking and roto.
- Check background faces in wide shots up close. When one region breaks, fix just that crop rather than regenerating the whole shot.
- Generate at or near your delivery resolution where you can. An HD to 4K upscale has to invent three of every four pixels, and one team tried 10 to 15 upscales on some shots before switching to CG.
- Don't expect one model to do everything. Keep a note of which model holds camera movement and which holds faces.
The single-shot trap
Most AI video tools are built to win the hero shot, but studio work is won on shot 200. By then you need the same colour pipeline as shot 1, a revision that only touches what the client flagged, and data that never leaks into another show.
Lee Jaeseon of CJ 4DPLEX's SX Studio put it in one line: "The success of one shot does not guarantee the success of hundreds of shots."
What CJ 4DPLEX found extending films to 270 degrees
CJ 4DPLEX, the team behind ScreenX, presented a production case study at the AI Content Festival in Seoul on 13 November. They tried commercial AI tools for extending films onto the side walls and walked away, for three reasons:
| Blocker | Why it matters at volume |
|---|---|
| No tool could outpaint beyond the frame above 8K | ScreenX delivery is far wider than a standard frame |
| Revisions and colour correction were hard to control | Every note re-rolls more than the note asked for |
| Weak on TPN (the Trusted Partner Network, Hollywood's content security audit) | Studio IP can't go into a shared tool |
So they're building their own platform, working title GEN.AI, using AI only where it works (mostly extending camera moves and character motion). Each studio's IP lives in its own workspace with its own permissions, which their CIO said Marvel and Disney asked for. A ScreenX version takes about four weeks with 10+ outside partners, so the goal is throughput, not a demo.



