mixed up beta schedules - linear and sqrt_linear

[eps696]

it seems beta schedules in comfyui are inconsistent with those from diffusers library.

in diffusers, e.g. euler (it's the same elsewhere):

        elif beta_schedule == "linear":
            self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
        elif beta_schedule == "scaled_linear":
            # this schedule is very specific to the latent diffusion model.
            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2

in comfyui:

    if schedule == "linear":
        betas = torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
. . .
    elif schedule == "sqrt_linear":
        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)

are the names simply mistaken or there is a logics behind such swap?

[comfyanonymous]

That ComfyUI code is from the original stable diffusion repo.

[eps696]

how does it change the fact that your beta schedulers are incorrect? [maybe it was originally a diffusers' mistake, but they seemeed to correct it now]

[comfyanonymous]

My beta schedulers are correct. That's what they call it in the original stable diffusion code where I copied it from.

[eps696]

glad to hear you're so confident about that. hope this issue will allow others to understand better how torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2 can be called linear, while betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) pretends to be sqrt_linear. you may close this issue if you want to.

that said, thanks a ton and a great respect for the coolest product in the area, it's really a fresh air amongst anything else.