StableMaterials: Enhancing Diversity in Material Generation via Semi-Supervised Learning
Comparison between LDM and LCM
We compare the generative performance of the distilled Latent Consistency model vs the original Latent Diffusion model. We show minimal loss in quality (if not better in some cases) with significantly improved inference speed. Each generation using the non-distilled LDM requires 50 + 20 diffusion steps (7.8s at 1K) compared to the 4 + 1 of the LCM (1.7s at 1K).
| Render | Basecolor | Normal | Height | Roughness | Metallic | |
|---|---|---|---|---|---|---|
|
|
||||||
"Brick wall with crumbling paint." |
||||||
|
LDM |
 |
|
|
 |
 |
|
|
LCM |
 |
|
|
 |
 |
|
|
|
||||||
"Rough stone wall with moss and lichen growing on it." |
||||||
|
LDM |
 |
|
|
 |
 |
|
|
LCM |
 |
|
|
 |
 |
|
|
|
||||||
"Rusty metal surface with peeling paint." |
||||||
|
LDM |
 |
|
|
 |
 |
|
|
LCM |
 |
|
|
 |
 |
|
|
|
||||||
"Bathroom tiles." |
||||||
|
LDM |
 |
|
|
 |
 |
|
|
LCM |
 |
|
|
 |
 |
|
|
|
||||||
"Old wooden floor with scratches and stains." |
||||||
|
LDM |
 |
|
|
 |
 |
|
|
LCM |
 |
|
|
 |
 |
|