Fine-grained, continuous control over multiple attributes at once (e.g., exact eye-openness and car width) on top of any pretrained diffusion model. A decoupled cross-attention layer that cleanly ...