One of the hardest problems in AI video generation is "subject consistency". You generate a cool astronaut in scene 1, but in scene 2, they have a different helmet and suit.
What is In-Context Generation?
Borrowing from LLM terminology, "In-Context" means providing examples (context) to the model at inference time. For UniVideo, this context comes in the form of Reference Images.
Instead of describing "a futuristic soldier" in text and hoping for the best, you upload an image of the specific soldier you want. UniVideo's MLLM "sees" this image, extracts the features (armor design, color palette, face), and uses them as a hard constraint for the generation process.
How to Use It
Upload Reference
In the UniVideo dashboard, select the "In-Context" tab. Drag and drop your character sheet nor reference frame.
Define the Action
Type your prompt: "The character is running through a forest." The model knows "The character" refers to your image.
Generate
UniVideo generates the motion while locking onto the visual identity of the reference.
Use Cases
-
Virtual Influencers
Maintain the same face and outfit across daily vlogs generated entirely by AI.
-
Storyboarding
Take a crude sketch and animate it into a polished animatic without changing the composition.