ChangeAngle V2 – KonText LoRA

In KonText, when prompting for a camera movement — like shifting left, right, or changing angle — the model typically rotates the main subject instead of moving the camera around it, leaving the background untouched.

This LoRA corrects that behavior by allowing you to virtually control the camera, generating a realistic change in viewpoint while preserving the subject’s pose, identity, and spatial coherence.

Use the prompt:
ChangeAngle, pan camera right

—or variants like—
ChangeAngle, pan camera left
ChangeAngle, tilt camera down
ChangeAngle, zoom-out
ChangeAngle, tilt camera top view
ChangeAngle, pan camera opposite rear view

You can also combine multiple movements for more cinematic control:
ChangeAngle, tilt camera slightly down, pan camera left, zoom-out

By default, this model actually shifts the virtual camera rather than rotating the subject in place.

✦ What’s new in this version:

• Added support for zoom-out, top view and opposite rear view

• Improved consistency and coherence across multiple generations

• More stable preservation of subject identity during angle changes

✦ Important notes:

• It sometimes confuses left and right, but the correct orientation is usually achieved after a few generations.

• On complex scenes, the background may be over-modified or altered.

Trained over 3000 steps on a dataset focused on subject rotation, angle variation, zooming, and cinematic framing.

✦ Ideal for:

• Building multi-angle datasets or before/after sequences

• Simulating camera motion across a visual sequence

• Ensuring visual continuity in storyboards or game/film concepts

• Generating in/out frames for video interpolation models like WAN