Stroke-based rendering is often optimized as a pixel-level reconstruction task. This over-reliance on global error reduction not only degrades strokes into fragmented curves due to a lack of structured planning, but also allows locally defective primitives to be concealed by overlapping layers. To address these issues, we introduce a model-guided framework for stroke-audited rendering. Our framework uses Quadratic Bezier Splatting to analytically model continuous curved strokes with efficient differentiable alpha compositing. Built on this renderer, a feedforward multi-scale stroke decoder provides a structural initialization with learned stroke-scale allocation, while test-time stroke optimization supplies the reconstruction capacity needed for high-fidelity refinement. We further propose a single-stroke audit mechanism that explicitly evaluates local stroke quality, preventing defective primitives from surviving optimization through layer-wise compensation. Compared with existing methods, our approach achieves coherent layouts, improves PSNR by up to 5.46 dB, yields a 16% more compact representation, and accelerates combined renderer forward and backward computation by a factor of 220.3.
Figures & tables
Figure 1: Motivation. Left: Image-blur-based coarse-to-fine optimization mainly performs low-level smoothing and produces fragmented stroke layouts, while our semantic initialization yields more coherent multi-scale strokes. Right: Optimization alone can eventually converge but suffers from early over-smoothing, whereas model prediction alone provides good early structure but is difficult to refine. Combining semantic initialization with optimization achieves both well-structured strokes and stable convergence.
Figure 2: Framework. The Input image is represented by quadratic Bézier strokes rendered through an analytic SDF-based differentiable splatting renderer. A ViT-based feedforward predictor decodes multi-scale stroke queries into coarse, mid, and fine stroke groups, yielding stroke parameters as a structural initialization trained with Lmodel . Starting from this initialization, test-time optimization refines the canvas with Lopt . In parallel, the stroke audit module measures each stroke’s purity, benefit density, and residual density, and triggers adaptive pruning, splitting, and respawning to reallocate stroke capacity toward unresolved structures.
Figure 3: Qualitative comparison with state-of-the-art methods. Compared with existing approaches, our method produces more coherent structural layouts and better organized stroke flows while preserving cleaner local details. Baseline methods often exhibit fragmented strokes, unstable boundaries, and texture-driven artifacts, especially under limited stroke budgets.
Dataset
DIV2K 600×600 / 5K
Gallery 600×800 / 6K
Efficiency
Paradigm
Method
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Time ↓
Size ↓
Search
Im2Oil Tong et al. (2022)
16.83
0.3602
0.4298
20.59
0.5056
0.4623
32.9ms
61.0K
Opt.
DPR Hu et al. (2023)
20.26
0.4614
0.4363
22.48
0.5501
0.4838
6.3ms
56.4K
FF
PT Liu et al. (2021)
14.40
0.2518
0.5084
17.22
0.4479
0.5325
3.1ms
18.0K
FF
L2P Huang et al. (2019)
20.72
0.5020
0.5023
24.86
0.6573
0.4068
16.8ms
31.4K
Opt.
SNP Zou et al. (2021)
14.79
0.3787
0.5601
17.23
0.4345
0.5936
276.1ms
96.2K
Table 1: Quantitative comparison on DIV2K and Gallery with 5K and 6K strokes, respectively. Time is average time per stroke; Size is the final stroke representation size.
Figure 4: Curve quality comparison. Polyline-based representations approximate curved strokes using piecewise-linear segments, while our analytic quadratic Bezier splatting directly represents continuous stroke trajectories, yielding strokes with coordinated structural semantics.
Method
#Prim.
Fwd(ms)
Bwd(ms)
Speedup
PSNR
DiffVG for Quadratic Bezier
2K
190.6
902.3
1.0 ×
27.91
Neural Render for Line
2K
312.4
1488.7
0.6 ×
25.84
Polyline Splatting
2K
4.1
16.8
52.2 ×
29.74
Quadratic Bezier, sampled (64 Gaussians/stroke)
128K
2.34
3.90
175.1 ×
28.96
Bezier Splatting, open Liu et al. (2025)
384K
4.81
5.05
110.8 ×
35.87
Bezier Splatting, closed Liu et al. (2025)
5.38M
15.76
13.81
37.0 ×
32.14
Table 2: Comparison with differentiable rendering methods under a 2K-stroke budget. #Prim. is the number of runtime primitives; speedup is relative to the total forward and backward time of DiffVG. Our implementation applies the renderer-level optimizations of Bezier Splatting Liu et al. (2025) (forward-state reuse, warp-level gradient reduction, and analytical tile culling); Bezier Splatting rows use its released configuration. Best and second best timing, speedup, and PSNR values are marked among the listed methods.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
Purity ↑
Benefit ↑
Residual ↓
Time ↓
Size ↓
Full (FF + Opt)
28.52
0.872
0.171
0.9949
1.5496
0.0630
26.9ms
21.2K
Initialization & Architecture
w/o Opt (FF Only)
15.65
0.3098
0.6028
0.7042
0.4488
0.3502
/
1.0K
w/o Opt (FF Only, 5K strokes)
20.58
0.5313
0.4067
0.9209
1.1297
0.0739
/
21.2K
Random Init (Opt Only)
28.20
0.8495
0.1867
0.9912
1.5381
0.0732
26.6ms
21.2K
Edge Init (Heuristic)
28.35
0.8480
0.1875
0.9924
1.5403
0.0743
26.9ms
21.2K
Table 3: Ablation study on structural initialization, multi-scale query design, and adaptive stroke auditing. “w/o Opt” denotes the raw output of the feedforward initializer before test-time optimization, using either its native stroke count or the same 5K-stroke budget as the full model. Missing or non-applicable metrics across different testing regimes are denoted by “/”.
Figure 5: Initialization comparison. Gaussian coarse-to-fine initialization mainly provides low-frequency smoothing and often produces fragmented stroke layouts after optimization. In contrast, our feedforward initialization generates more coherent structural organization, leading to fewer dead strokes, faster convergence, and better final reconstruction quality.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Stroke quality comparison. Comparison of stroke auditing metrics on DIV2K( 600×600 ). Our framework establishes a significant margin over all baselines, demonstrating that it generates locally constructive and structurally meaningful strokes rather than unstructured pixel accumulations.
Group
Ratio
Curve length
Width
Role
Coarse
20%
0.24–0.56
0.12–0.30
large color regions
Mid
30%
0.08–0.22
0.04–0.12
medium-scale boundaries
Fine
50%
0.008–0.06
0.006–0.03
local residual hints
Appendix
Table 4: Stage-specific geometric priors for stroke templates normalized to local patch coordinates.
Figure 7: Patch-boundary behavior before and after global refinement. Columns show the target, patch-wise stitching, global rendering of the same initialized strokes, and the result after 5,000 global optimization steps. The bottom row enlarges the patch junction. Patches are 256×256 on a 1024×1024 base canvas.
Figure 8: Sensitivity to fixed within-layer ordering. Columns show the target and results under original, random, and reversed within-layer orders. Full images appear above, with the same region enlarged below. All conditions start from identical stroke geometry, colors, and opacities, differing only in rendering order, and use the same optimization budget. Within-layer reordering has little effect on reconstruction quality.
Order
PSNR ↑
SSIM ↑
LPIPS ↓
Original
31.28
0.8704
0.1731
Random
31.35
0.8709
0.1702
Reversed
31.35
0.8697
0.1748
Appendix
Table 5: Within-layer stroke-order ablation. Mean results on Im2Oil Gallery under the 600×600 , 5,000-stroke setting. Every condition uses the reconstruction-selected output exported by the pipeline.
Figure 9: Quadratic Bézier and line-segment primitives. Columns show the target, quadratic-stroke reconstruction, and line-segment reconstruction. Both variants use the same stroke capacity and optimization budget.
Figure 10: Enlarged qualitative results and oil rendering. Columns show the input, plain rendering, and oil rendering. Both rendering variants share identical fitted stroke geometry and colors; enlarged outputs are rerasterizations without additional fitting.
Dataset
Use
License / terms
Source
Flickr2K
Training
No unified named dataset license is specified in the original release; images were collected from Flickr. We use the data for non-commercial academic training only and do not redistribute the images.
Official usage terms restrict the dataset to academic research purposes; image copyrights remain with the original owners.
Official DIV2K page: https://data.vision.ee.ethz.ch/cvl/DIV2K/
Im2Oil Gallery [ 33 ]
Evaluation
The official Im2Oil project repository is released under the CC0-1.0 license.
Official repository: https://github.com/TZYSJTU/Im2Oil
Appendix
Table 6: Dataset licenses or usage terms for the assets used in this paper.
Method
Struct. Plan.↑
Artistic Qual.↑
Human-like↑
Stroke Qual.↑
SNP [ 38 ]
2.41
2.38
2.29
2.35
PT [ 27 ]
2.78
2.85
2.71
2.74
Im2Oil [ 33 ]
3.15
3.22
3.08
3.11
DPR [ 15 ]
3.33
3.19
3.14
3.08
L2P [ 17 ]
3.52
3.58
3.47
3.44
DP [ 26 ]
3.91
3.97
3.85
3.89
Appendix
Table 7: User study results on perceptual painting quality. Participants rated each result on a 1–5 Likert scale; higher scores indicate better perceptual quality.