🦴SKDream: Controllable Multi-view and 3D Generation with Arbitrary Skeletons

Text prompts can describe visual concepts, but how can we describe shapes?


Overview

- Skeletons can effectively represent and describe object anatomy and pose. Unfortunately, past studies are often limited to human skeletons. In this work, we generalize skeletal conditioned generation to arbitrary structures.
- First, we design a reliable mesh skeletonization pipeline to generate a large-scale mesh-skeleton paired dataset. Based on the dataset, a multi-view and 3D generation pipeline is built. Our method achieves instant generation of multi-view and 3D contents which are aligned with given skeletons.


Generation Results on Synthetic Skeletons

Our work achieves conditional generation with arbitrary skeletal structures. Multi-view images are first generated, and 3D meshes are reconstructed and refined. The generated contents are aligned with the given skeletons. The skeletons below are from our Objaverse-SK val set.

A creature with a pig head and a fish body. A carp, elongated body, large golden scales. A rattle snake.
An alien, with a large head, big eyes. Cartoon character in Fall Guys. A robot.
A duck. A plesiosaur with a long neck. A swan with feathers, long neck.
A seagull, white feathers with gray on the back and wings. A small bat with leathery wings, pointed ears, dark fur. A penguin, black back, white belly, flipper-like wings.
A polar bear, large body, black nose. A wolf, thick grey fur, pointed ears, bushy tail. A bull, muscular body, short black fur, curved horns.
A Golden Retriever, friendly expression, floppy ears. A donkey, grey fur, long ears. A sea turtle, broad oval shell, smooth olive-green scales.
A scorpion, segmented body, curved tail with stinger, two large pincers. A black and red spider. A lizard, elongated body, scaly skin.
A potted plant. A green cactus, with multiple arms. A tree stump, roots and trunk.

Generation Results on Human-made Skeletons

Our method can generalize to arbitrary human-made skeletons (from ModelResource dataset), enabling flexible and efficient structural description. Disscussion can be found in Appendix.

French Bulldog. Dog resembling a wolf. Rat and lizard.
Triceratops dinosaur. Crocodile, turtle, bird, and camel figures. Fish resembling a bird.
Nintendo character resembling a spiky-headed creature. Pokemon resembling a fox and beaver. Bunny.
Mouse with a tail. Mouse-like teddy bear with outstretched arms. Turtle Pokemon Baby.
Horned Pokemon character. Cat with a bow tie and tail. Small teddy bear styled as a Nintendo Pokemon character.

Animation Results

Our method can be applied to produce animation. Given a motion sequence of a skeleton, the corresponding mesh is generated and rigged with the skeleton. Then the mesh can be animated by the skeleton.

A squid, long head, tentacles. A duck.
A kitten, cute face, long tail. A zebra.