Image Paper Image arXiv Image Video Image GitHub (TBD)
Image

Abstract

Human Cinemegraph Generation

Image Image

Methodology

We propose solving this problem in normal space. By doing so, it is possible to tackle the problem without real data.

Pipeline

Image

Problem Decomposition

Here we detail how we decompose the problem to approach it through the normal space.

Stage 1: Normal estimation
  • Image
  • Image
  • Image
  • Image
Stage 2: Cyclic and Controllable Animation (Sec 3.1)

Wind Cyclic UNet

Image
Stage 3: Normal Guided Synthesis (Sec. 3.2)
Image Input
Image Normal map
Animation
Shading
Image Reflectance
RGB

Synthetic Data

We train our Cyclic and Controllable UNet using only synthetic data.

Comparison

We compare results obtained with a state-of-the-art method and variations of our approach.

  • Controllable Animation [27]
    RGB
    Normals
    Estimated Normals
  • Controllable Animation [27]
    RGB
    Normals
    Estimated Normals
  • Controllable Animation [27]
    RGB
    Normals
    Estimated Normals

Wind Control

We illustrate here the possibility of controlling the generated video with the wind direction.

Qualitative results

More results

Failures

BibTeX

Image 
[TBD]