Outpainting with Diffusion Models

Module: Generative AI | Difficulty: Advanced

Outpainting Problem

Extend image beyond boundaries while maintaining coherence.

Boundary Conditioning

Use the boundary region as conditioning.

Autoregressive Outpainting

1. Generate left/right/top/bottom strip
2. Use generated strip as new boundary
3. Repeat

Coherence via Attention

def outpaint(model, image, n_steps=4, direction='right'):
    result = image.clone()
    for _ in range(n_steps):
        h, w = result.shape[2:]
        if direction == 'right':
            canvas = torch.zeros(1, 3, h, w*2)
            canvas[:, :, :, :w] = result
            mask = torch.zeros(1, 1, h, w*2)
            mask[:, :, :, :w] = 1
        # Inpaint the blank region
        result = inpaint_step(model, canvas, mask, result, t=500)
    return result

Research Insight: Outpainting works best when the image has strong structure (e.g., landscapes, architecture). Random textures (e.g., grass, clouds) are easier than structured content (e.g., faces, text).