ControlNet
ControlNet: An architecture that adds spatial conditions to a diffusion model without replacing its core weights. A trainable copy of the encoder receives maps such as edges or poses and combines them with the frozen model. It enables more direct structural control than a prompt alone.
In italiano: ControlNetLegge II · Probabilità, softmax e modelli generativiA trainable copy of the encoder receives maps such as edges or poses and combines them with the frozen model. It enables more direct structural control than a prompt alone.
Examples
- Canny edges
- Depth map
- Human pose