Latent Consistency Models

A latent diffusion generator works in a learned representation rather than directly carrying out every denoising operation in pixel space. The associated autoencoder connects image pixels with that representation. Latent Consistency Models builds on consistency learning to reduce the sampling steps needed under its studied configurations. Its mapping predicts a near-clean endpoint from points along an augmented probability-flow ODE trajectory in latent space. Consistency asks points on the same trajectory to map to the same endpoint; it does not mean unrelated prompts should produce the same image. At the endpoint time epsilon, the function is constrained to return the supplied endpoint value, f(x, epsilon) equal to x. This boundary condition and trajectory invariance describe different properties. The original LCM procedure distills from a pretrained classifier-free-guided diffusion teacher. This is a particular training setup: the broader Consistency Models paper also investigates training without a pretrained teacher, so that alternative should not be confused with the cited LCM distillation. LCM-LoRA packages learned low-rank acceleration parameters for the studied Stable Diffusion families and compatible fine-tunes. Its report supports transfer in those configurations, rather than a guarantee that any adapter works with every checkpoint. Adversarial Diffusion Distillation is a distinct related few-step approach using an adversarial loss with a teacher signal. Across these methods, fewer network evaluations can reduce one part of generation cost, but neither step count nor a paper’s title proves near-zero latency on a phone. The decoder, model size, resolution, guidance, and implementation remain part of the measured image-generation system.