Pixal3D Multiview → GLB

Upload one calibrated ZIP containing transforms.json at its root and 4–8 PNG/JPG/JPEG/WEBP images. This is real multiview conditioning using Pixal3D's *_mv weights—there is no text-to-image step.

Every frames[*].file_path must map to exactly one image in the ZIP. Frame 1 must explicitly set is_canonical_front: true; every frame supplies a Blender/NeRF 4×4 camera-to-world matrix and real FOV in radians. The pipeline uses each supplied camera pose and distance; it never invents a pose.

Detailed face, eye, and hair images are allowed only as additional, truly calibrated frames in the same object coordinate system. A crop-only or guessed-pose close-up is not valid input and can harm reconstruction.