NeRF for Image Generation

Module: Generative AI | Difficulty: Advanced

3D-Aware GAN

where are camera parameters.

GRAF

GIRAFFE

Volume Rendering

def volume_render(density, color, t_values):
    delta = t_values[:, 1:] - t_values[:, :-1]
    alpha = 1 - torch.exp(-density * delta)
    T = torch.cumprod(1 - alpha + 1e-10, dim=-1)
    T = torch.cat([torch.ones_like(T[:,:1]), T[:,:-1]], dim=-1)
    weights = T * alpha
    return (weights * color).sum(dim=-1)

Research Insight: GIRAFFE enables compositional generation — you can change objects, backgrounds, and camera viewpoints independently. This is impossible with 2D GANs.