Table 5 Performance comparison of different rendering models in ethnic painting style transfer tasks.
From: A painting art rendering system by deep learning framework and machine translation
Model | FID | LPIPS | StyleSimilarity | Generation speed (s/img) | Key limitations |
|---|---|---|---|---|---|
CNN-Style | 45.3 | 0.28 | 80.0% | 0.22 | Texture breaks; noticeable discontinuity in Miao silver “sawtooth pattern” |
VGG-19 | 41.6 | 0.25 | 82.0% | 0.25 | Harsh color transitions; blurred layering in Tibetan Thangka mineral pigments |
Vanilla GAN | 37.8 | 0.21 | 85.0% | 0.30 | Mode collapse; repeated “color-layered” generation in Yi lacquerware |
StyleGAN3 | 32.5 | 0.18 | 86.5% | 0.85 | Blurred boundaries between gold leaf base and mineral colors in Thangka |
CUT | 35.8 | 0.21 | 84.2% | 0.78 | Local distortion in Miao silver “symmetrical patterns” |
Diffusion Model | 30.1 | 0.16 | 88.3% | 1.20 | Insufficient detail restoration for Yi lacquerware “floral painting”; low generation efficiency |
IGAN | 28.2 | 0.12 | 91.0% | 0.35 | No obvious defects; high fidelity in technique and cultural feature reproduction |
