Autoguidance
A guidance technique originating from image diffusion models that uses a weakened version of the same model—such as an under-trained checkpoint or a smaller variant—as a negative reference. Generation quality is amplified by contrasting predictions from the strong and weak models, though each inference step requires a forward pass through both models, doubling computational cost.
Timeline (last 90 days)
Autoguidance is a guidance technique originating from image diffusion models that uses a weakened version of the same model as a negative reference, amplifying generation quality by contrasting the prediction differences between the strong and weak models.
Autoguidance requires separate forward passes for both the strong and weak models at each inference step, doubling the computational cost.
Classifier-free guidance (CFG) uses unconditional predictions as the weak signal, whereas autoguidance uses a weak model as the weak signal.
All Facts (3)
Autoguidance is a guidance technique originating from image diffusion models that uses a weakened version of the same model as a negative reference, amplifying generation quality by contrasting the prediction differences between the strong and weak models.
50%UnverifiedAutoguidance requires separate forward passes for both the strong and weak models at each inference step, doubling the computational cost.
50%UnverifiedClassifier-free guidance (CFG) uses unconditional predictions as the weak signal, whereas autoguidance uses a weak model as the weak signal.
50%