openai / consistency_models

Hi,
thank you for your great work on this project!
I have been comparing the noise schedule implementation between this project and EDM, and I noticed some odd differences.
In particular, I was looking at this code snippet:

consistency_models/cm/script_util.py

Lines 94 to 101 in edfe91e

    
           diffusion = KarrasDenoiser( 
        
               sigma_data=0.5, 
        
               sigma_max=sigma_max, 
        
               sigma_min=sigma_min, 
        
               distillation=distillation, 
        
               weight_schedule=weight_schedule, 
        
           ) 
        
           return model, diffusion

The image tensors in both projects are normalized in the range [-1, 1]. Based on my understanding, the parameter "sigma_data" should be set to 1.0, similar to the EDM settings. Could you kindly clarify the reason behind the difference in the "sigma_data" value between the two projects?

	diffusion = KarrasDenoiser(
	sigma_data=0.5,
	sigma_max=sigma_max,
	sigma_min=sigma_min,
	distillation=distillation,
	weight_schedule=weight_schedule,
	)
	return model, diffusion

The parameter sigma_data for KarrasDenoiser.