Parallel Controllable Texture Synthesis

Sylvain Lefebvre, Hugues Hoppe

SIGGRAPH 2005

24(3), 777-786

Parallel Controllable Texture Synthesis

2

Outline

• Introduction• Parallel synthesis method• Synthesis control• Result

Parallel Controllable Texture Synthesis

3

Introduction

• Sample-based texture synthesis analyzes a given exemplar to create visually similar images.– tiling methods are the fastest – patch optimization methods produce some of the best results– neighborhood-matching algorithms allow greater fine-scale ada

ptability

• Our interest is in applying synthesis to define infinite, aperiodic, deterministic content from a compact representation. we present a new neighborhood-matching method.

Parallel Controllable Texture Synthesis

4

Parallel synthesis method –Basic scheme

• E[u]=E[S[p]]– E: m×m exemplar image– S: synthesized image– p Z∈ 2

– u Z∈ 2

• image pyramid

S0, S1, …, (SL=S) in coarse-to-fine order, where L=log2m.

(Figure 2)

Parallel Controllable Texture Synthesis

5

Parallel synthesis method –steps

Parallel Controllable Texture Synthesis

6

Parallel synthesis method – Upsampling & Jitter

– hl regular output spacing : 1 for pyramid, 2L-1 for a stack

– Hash Function H: Z2 -> [-1,+1]2

– per-level randomness parameter rl : [0,1]

Parallel Controllable Texture Synthesis

7

Parallel synthesis method –Correction

• For each pixel p, we gather the pixel colors of its 5×5 neighbor-hood at the current level, represented as a vector NSl(p) . This neighborhood is compared with exemplar neighborhoods NEl(u) to find the L2 best matching one.

Parallel Controllable Texture Synthesis

8

Parallel synthesis method – Gaussian image stack

• traditional Gaussian image pyramid often results in synt

hesized features that align with a coarser “grid”.

• because ancestor coordinates in the synthesis pyramid are snapped to the quantized positions of the exemplar pyramid.

Parallel Controllable Texture Synthesis

9

Parallel synthesis method – Gaussian image stack

• augment the exemplar image on all sides to have size 2m×2m:– an actual larger texture– a tiling if the exemplar is toro

idal– reflected copies of the exem

plar

• Gaussian filtering each level, without subsampling.

• Reassign hl = 2L-l.

Synthesis control –Multiscale randomness control

• The randomness parameters r

l set the jitter amplitude at each level, and thus provide a form of “spectral variation control” .

Parallel Controllable Texture Synthesis

11

Synthesis control –Spatial modulation over source

• preserving the integrity of selected texture elements in nonstationary textures.

Parallel Controllable Texture Synthesis

12

Synthesis control –Spatial modulation over output

• to roughen a surface pattern in areas of wear or damage.

Synthesis control –Feature drag-and-drop

• Sl[p]:=((uF)+(p-pF)) mod m

if ||p-pF|| < rF

– pF : the desired exemplar coordinate

– rF : exemplar radius

– uF : circle center

• rF = ril/L + ro(L-l)/L

– ri : inner radius

– ro : outer radius

Parallel Controllable Texture Synthesis

14

Result

• additional results