Studio Engine
Generation that respects the math of the model.
Most AI video tools fail silently because they hand the model dimensions and frame counts it cannot represent. The Studio Engine normalises every workflow before dispatch, so a render either runs correctly or tells you precisely why it cannot.
Photography
Text-to-image and reference-image Face-Lock injection, with your creator’s LoRA triggers and BrandBrain aesthetic rules merged into every prompt automatically. Batch between one and ten variations from a single configured prompt.
- Face-Lock. A reference still constrains identity so your creator stays recognisable across a thousand frames.
- Automatic LoRA triggers. Character activation tokens are injected without you retyping them, at a tuned strength of 0.85.
- BrandBrain inheritance. Palette, tone, and forbidden words are applied before dispatch, not after review.
Cinematic Video
Image-to-video rendering with a workflow normaliser that inserts an ImageScale node immediately before the VAE encode, forcing the input still to the exact canvas target. This eliminates the latent-dimension mismatch that otherwise produces silent, unusable output.
Canvas targets
- Landscape 16:9960×544
- Vertical Reel 9:16544×960
- Square 1:1768×768
Duration modes · strict 8n + 1 frame math
- 5.4s video mode129 frames @ 24 FPS
- 10.7s video mode257 frames @ 24 FPS
Why your renders stop crashing with Unexpected token ‘<’
When a GPU worker is overloaded or restarting, it does not return JSON — it returns an HTML error page from a reverse proxy. Naively calling .json() on that response throws a parse error that surfaces to the user as an incomprehensible crash, usually after they have already waited through a long render.
Every HTTP and WebSocket call in the Studio Engine validates response.ok and the Content-Type header before parsing. Non-JSON responses are intercepted and converted into structured errors that name the actual failure:
{
"error": "Worker Unreachable (502/504)",
"statusCode": 502
}