Multi Layer Image Composition And Overlay Blending

via “image mixing with multi-image concept blending”

Kandinsky 2 — multilingual text2image latent diffusion model

Unique: Operates in CLIP embedding space rather than pixel or latent space, enabling semantic blending of image concepts. Uses diffusion prior to map interpolated embeddings back to coherent images, allowing fine-grained control over blend ratios without retraining.

vs others: Provides explicit control over image blending weights and text guidance, unlike simple image averaging or GAN-based morphing, and leverages the diffusion prior for higher-quality outputs than direct embedding interpolation.