Upload 2 files

dpm2mv2/sampling.py

@@ -0,0 +1,687 @@
+import math
+
+from scipy import integrate
+import torch
+from torch import nn
+from torchdiffeq import odeint
+import torchsde
+from tqdm.auto import trange, tqdm
+
+from . import utils
+
+
+def append_zero(x):
+    return torch.cat([x, x.new_zeros([1])])
+
+
+def get_sigmas_karras(n, sigma_min, sigma_max, rho=7., device='cpu'):
+    """Constructs the noise schedule of Karras et al. (2022)."""
+    ramp = torch.linspace(0, 1, n)
+    min_inv_rho = sigma_min ** (1 / rho)
+    max_inv_rho = sigma_max ** (1 / rho)
+    sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+    return append_zero(sigmas).to(device)
+
+
+def get_sigmas_exponential(n, sigma_min, sigma_max, device='cpu'):
+    """Constructs an exponential noise schedule."""
+    sigmas = torch.linspace(math.log(sigma_max), math.log(sigma_min), n, device=device).exp()
+    return append_zero(sigmas)
+
+
+def get_sigmas_polyexponential(n, sigma_min, sigma_max, rho=1., device='cpu'):
+    """Constructs an polynomial in log sigma noise schedule."""
+    ramp = torch.linspace(1, 0, n, device=device) ** rho
+    sigmas = torch.exp(ramp * (math.log(sigma_max) - math.log(sigma_min)) + math.log(sigma_min))
+    return append_zero(sigmas)
+
+
+def get_sigmas_vp(n, beta_d=19.9, beta_min=0.1, eps_s=1e-3, device='cpu'):
+    """Constructs a continuous VP noise schedule."""
+    t = torch.linspace(1, eps_s, n, device=device)
+    sigmas = torch.sqrt(torch.exp(beta_d * t ** 2 / 2 + beta_min * t) - 1)
+    return append_zero(sigmas)
+
+
+def to_d(x, sigma, denoised):
+    """Converts a denoiser output to a Karras ODE derivative."""
+    return (x - denoised) / utils.append_dims(sigma, x.ndim)
+
+
+def get_ancestral_step(sigma_from, sigma_to, eta=1.):
+    """Calculates the noise level (sigma_down) to step down to and the amount
+    of noise to add (sigma_up) when doing an ancestral sampling step."""
+    if not eta:
+        return sigma_to, 0.
+    sigma_up = min(sigma_to, eta * (sigma_to ** 2 * (sigma_from ** 2 - sigma_to ** 2) / sigma_from ** 2) ** 0.5)
+    sigma_down = (sigma_to ** 2 - sigma_up ** 2) ** 0.5
+    return sigma_down, sigma_up
+
+
+def default_noise_sampler(x):
+    return lambda sigma, sigma_next: torch.randn_like(x)
+
+
+class BatchedBrownianTree:
+    """A wrapper around torchsde.BrownianTree that enables batches of entropy."""
+
+    def __init__(self, x, t0, t1, seed=None, **kwargs):
+        t0, t1, self.sign = self.sort(t0, t1)
+        w0 = kwargs.get('w0', torch.zeros_like(x))
+        if seed is None:
+            seed = torch.randint(0, 2 ** 63 - 1, []).item()
+        self.batched = True
+        try:
+            assert len(seed) == x.shape[0]
+            w0 = w0[0]
+        except TypeError:
+            seed = [seed]
+            self.batched = False
+        self.trees = [torchsde.BrownianTree(t0, w0, t1, entropy=s, **kwargs) for s in seed]
+
+    @staticmethod
+    def sort(a, b):
+        return (a, b, 1) if a < b else (b, a, -1)
+
+    def __call__(self, t0, t1):
+        t0, t1, sign = self.sort(t0, t1)
+        w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign)
+        return w if self.batched else w[0]
+
+
+class BrownianTreeNoiseSampler:
+    """A noise sampler backed by a torchsde.BrownianTree.
+
+    Args:
+        x (Tensor): The tensor whose shape, device and dtype to use to generate
+            random samples.
+        sigma_min (float): The low end of the valid interval.
+        sigma_max (float): The high end of the valid interval.
+        seed (int or List[int]): The random seed. If a list of seeds is
+            supplied instead of a single integer, then the noise sampler will
+            use one BrownianTree per batch item, each with its own seed.
+        transform (callable): A function that maps sigma to the sampler's
+            internal timestep.
+    """
+
+    def __init__(self, x, sigma_min, sigma_max, seed=None, transform=lambda x: x):
+        self.transform = transform
+        t0, t1 = self.transform(torch.as_tensor(sigma_min)), self.transform(torch.as_tensor(sigma_max))
+        self.tree = BatchedBrownianTree(x, t0, t1, seed)
+
+    def __call__(self, sigma, sigma_next):
+        t0, t1 = self.transform(torch.as_tensor(sigma)), self.transform(torch.as_tensor(sigma_next))
+        return self.tree(t0, t1) / (t1 - t0).abs().sqrt()
+
+
+@torch.no_grad()
+def sample_euler(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
+    """Implements Algorithm 2 (Euler steps) from Karras et al. (2022)."""
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    for i in trange(len(sigmas) - 1, disable=disable):
+        gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
+        eps = torch.randn_like(x) * s_noise
+        sigma_hat = sigmas[i] * (gamma + 1)
+        if gamma > 0:
+            x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
+        denoised = model(x, sigma_hat * s_in, **extra_args)
+        d = to_d(x, sigma_hat, denoised)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
+        dt = sigmas[i + 1] - sigma_hat
+        # Euler method
+        x = x + d * dt
+    return x
+
+
+@torch.no_grad()
+def sample_euler_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    """Ancestral sampling with Euler method steps."""
+    extra_args = {} if extra_args is None else extra_args
+    noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        d = to_d(x, sigmas[i], denoised)
+        # Euler method
+        dt = sigma_down - sigmas[i]
+        x = x + d * dt
+        if sigmas[i + 1] > 0:
+            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
+    return x
+
+
+@torch.no_grad()
+def sample_heun(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
+    """Implements Algorithm 2 (Heun steps) from Karras et al. (2022)."""
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    for i in trange(len(sigmas) - 1, disable=disable):
+        gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
+        eps = torch.randn_like(x) * s_noise
+        sigma_hat = sigmas[i] * (gamma + 1)
+        if gamma > 0:
+            x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
+        denoised = model(x, sigma_hat * s_in, **extra_args)
+        d = to_d(x, sigma_hat, denoised)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
+        dt = sigmas[i + 1] - sigma_hat
+        if sigmas[i + 1] == 0:
+            # Euler method
+            x = x + d * dt
+        else:
+            # Heun's method
+            x_2 = x + d * dt
+            denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
+            d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
+            d_prime = (d + d_2) / 2
+            x = x + d_prime * dt
+    return x
+
+
+@torch.no_grad()
+def sample_dpm_2(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
+    """A sampler inspired by DPM-Solver-2 and Algorithm 2 from Karras et al. (2022)."""
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    for i in trange(len(sigmas) - 1, disable=disable):
+        gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
+        eps = torch.randn_like(x) * s_noise
+        sigma_hat = sigmas[i] * (gamma + 1)
+        if gamma > 0:
+            x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
+        denoised = model(x, sigma_hat * s_in, **extra_args)
+        d = to_d(x, sigma_hat, denoised)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            # Euler method
+            dt = sigmas[i + 1] - sigma_hat
+            x = x + d * dt
+        else:
+            # DPM-Solver-2
+            sigma_mid = sigma_hat.log().lerp(sigmas[i + 1].log(), 0.5).exp()
+            dt_1 = sigma_mid - sigma_hat
+            dt_2 = sigmas[i + 1] - sigma_hat
+            x_2 = x + d * dt_1
+            denoised_2 = model(x_2, sigma_mid * s_in, **extra_args)
+            d_2 = to_d(x_2, sigma_mid, denoised_2)
+            x = x + d_2 * dt_2
+    return x
+
+
+@torch.no_grad()
+def sample_dpm_2_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    """Ancestral sampling with DPM-Solver second-order steps."""
+    extra_args = {} if extra_args is None else extra_args
+    noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        d = to_d(x, sigmas[i], denoised)
+        if sigma_down == 0:
+            # Euler method
+            dt = sigma_down - sigmas[i]
+            x = x + d * dt
+        else:
+            # DPM-Solver-2
+            sigma_mid = sigmas[i].log().lerp(sigma_down.log(), 0.5).exp()
+            dt_1 = sigma_mid - sigmas[i]
+            dt_2 = sigma_down - sigmas[i]
+            x_2 = x + d * dt_1
+            denoised_2 = model(x_2, sigma_mid * s_in, **extra_args)
+            d_2 = to_d(x_2, sigma_mid, denoised_2)
+            x = x + d_2 * dt_2
+            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
+    return x
+
+
+def linear_multistep_coeff(order, t, i, j):
+    if order - 1 > i:
+        raise ValueError(f'Order {order} too high for step {i}')
+    def fn(tau):
+        prod = 1.
+        for k in range(order):
+            if j == k:
+                continue
+            prod *= (tau - t[i - k]) / (t[i - j] - t[i - k])
+        return prod
+    return integrate.quad(fn, t[i], t[i + 1], epsrel=1e-4)[0]
+
+
+@torch.no_grad()
+def sample_lms(model, x, sigmas, extra_args=None, callback=None, disable=None, order=4):
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    sigmas_cpu = sigmas.detach().cpu().numpy()
+    ds = []
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        d = to_d(x, sigmas[i], denoised)
+        ds.append(d)
+        if len(ds) > order:
+            ds.pop(0)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        cur_order = min(i + 1, order)
+        coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)]
+        x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
+    return x
+
+
+@torch.no_grad()
+def log_likelihood(model, x, sigma_min, sigma_max, extra_args=None, atol=1e-4, rtol=1e-4):
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    v = torch.randint_like(x, 2) * 2 - 1
+    fevals = 0
+    def ode_fn(sigma, x):
+        nonlocal fevals
+        with torch.enable_grad():
+            x = x[0].detach().requires_grad_()
+            denoised = model(x, sigma * s_in, **extra_args)
+            d = to_d(x, sigma, denoised)
+            fevals += 1
+            grad = torch.autograd.grad((d * v).sum(), x)[0]
+            d_ll = (v * grad).flatten(1).sum(1)
+        return d.detach(), d_ll
+    x_min = x, x.new_zeros([x.shape[0]])
+    t = x.new_tensor([sigma_min, sigma_max])
+    sol = odeint(ode_fn, x_min, t, atol=atol, rtol=rtol, method='dopri5')
+    latent, delta_ll = sol[0][-1], sol[1][-1]
+    ll_prior = torch.distributions.Normal(0, sigma_max).log_prob(latent).flatten(1).sum(1)
+    return ll_prior + delta_ll, {'fevals': fevals}
+
+
+class PIDStepSizeController:
+    """A PID controller for ODE adaptive step size control."""
+    def __init__(self, h, pcoeff, icoeff, dcoeff, order=1, accept_safety=0.81, eps=1e-8):
+        self.h = h
+        self.b1 = (pcoeff + icoeff + dcoeff) / order
+        self.b2 = -(pcoeff + 2 * dcoeff) / order
+        self.b3 = dcoeff / order
+        self.accept_safety = accept_safety
+        self.eps = eps
+        self.errs = []
+
+    def limiter(self, x):
+        return 1 + math.atan(x - 1)
+
+    def propose_step(self, error):
+        inv_error = 1 / (float(error) + self.eps)
+        if not self.errs:
+            self.errs = [inv_error, inv_error, inv_error]
+        self.errs[0] = inv_error
+        factor = self.errs[0] ** self.b1 * self.errs[1] ** self.b2 * self.errs[2] ** self.b3
+        factor = self.limiter(factor)
+        accept = factor >= self.accept_safety
+        if accept:
+            self.errs[2] = self.errs[1]
+            self.errs[1] = self.errs[0]
+        self.h *= factor
+        return accept
+
+
+class DPMSolver(nn.Module):
+    """DPM-Solver. See https://arxiv.org/abs/2206.00927."""
+
+    def __init__(self, model, extra_args=None, eps_callback=None, info_callback=None):
+        super().__init__()
+        self.model = model
+        self.extra_args = {} if extra_args is None else extra_args
+        self.eps_callback = eps_callback
+        self.info_callback = info_callback
+
+    def t(self, sigma):
+        return -sigma.log()
+
+    def sigma(self, t):
+        return t.neg().exp()
+
+    def eps(self, eps_cache, key, x, t, *args, **kwargs):
+        if key in eps_cache:
+            return eps_cache[key], eps_cache
+        sigma = self.sigma(t) * x.new_ones([x.shape[0]])
+        eps = (x - self.model(x, sigma, *args, **self.extra_args, **kwargs)) / self.sigma(t)
+        if self.eps_callback is not None:
+            self.eps_callback()
+        return eps, {key: eps, **eps_cache}
+
+    def dpm_solver_1_step(self, x, t, t_next, eps_cache=None):
+        eps_cache = {} if eps_cache is None else eps_cache
+        h = t_next - t
+        eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
+        x_1 = x - self.sigma(t_next) * h.expm1() * eps
+        return x_1, eps_cache
+
+    def dpm_solver_2_step(self, x, t, t_next, r1=1 / 2, eps_cache=None):
+        eps_cache = {} if eps_cache is None else eps_cache
+        h = t_next - t
+        eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
+        s1 = t + r1 * h
+        u1 = x - self.sigma(s1) * (r1 * h).expm1() * eps
+        eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', u1, s1)
+        x_2 = x - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / (2 * r1) * h.expm1() * (eps_r1 - eps)
+        return x_2, eps_cache
+
+    def dpm_solver_3_step(self, x, t, t_next, r1=1 / 3, r2=2 / 3, eps_cache=None):
+        eps_cache = {} if eps_cache is None else eps_cache
+        h = t_next - t
+        eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
+        s1 = t + r1 * h
+        s2 = t + r2 * h
+        u1 = x - self.sigma(s1) * (r1 * h).expm1() * eps
+        eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', u1, s1)
+        u2 = x - self.sigma(s2) * (r2 * h).expm1() * eps - self.sigma(s2) * (r2 / r1) * ((r2 * h).expm1() / (r2 * h) - 1) * (eps_r1 - eps)
+        eps_r2, eps_cache = self.eps(eps_cache, 'eps_r2', u2, s2)
+        x_3 = x - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / r2 * (h.expm1() / h - 1) * (eps_r2 - eps)
+        return x_3, eps_cache
+
+    def dpm_solver_fast(self, x, t_start, t_end, nfe, eta=0., s_noise=1., noise_sampler=None):
+        noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+        if not t_end > t_start and eta:
+            raise ValueError('eta must be 0 for reverse sampling')
+
+        m = math.floor(nfe / 3) + 1
+        ts = torch.linspace(t_start, t_end, m + 1, device=x.device)
+
+        if nfe % 3 == 0:
+            orders = [3] * (m - 2) + [2, 1]
+        else:
+            orders = [3] * (m - 1) + [nfe % 3]
+
+        for i in range(len(orders)):
+            eps_cache = {}
+            t, t_next = ts[i], ts[i + 1]
+            if eta:
+                sd, su = get_ancestral_step(self.sigma(t), self.sigma(t_next), eta)
+                t_next_ = torch.minimum(t_end, self.t(sd))
+                su = (self.sigma(t_next) ** 2 - self.sigma(t_next_) ** 2) ** 0.5
+            else:
+                t_next_, su = t_next, 0.
+
+            eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
+            denoised = x - self.sigma(t) * eps
+            if self.info_callback is not None:
+                self.info_callback({'x': x, 'i': i, 't': ts[i], 't_up': t, 'denoised': denoised})
+
+            if orders[i] == 1:
+                x, eps_cache = self.dpm_solver_1_step(x, t, t_next_, eps_cache=eps_cache)
+            elif orders[i] == 2:
+                x, eps_cache = self.dpm_solver_2_step(x, t, t_next_, eps_cache=eps_cache)
+            else:
+                x, eps_cache = self.dpm_solver_3_step(x, t, t_next_, eps_cache=eps_cache)
+
+            x = x + su * s_noise * noise_sampler(self.sigma(t), self.sigma(t_next))
+
+        return x
+
+    def dpm_solver_adaptive(self, x, t_start, t_end, order=3, rtol=0.05, atol=0.0078, h_init=0.05, pcoeff=0., icoeff=1., dcoeff=0., accept_safety=0.81, eta=0., s_noise=1., noise_sampler=None):
+        noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+        if order not in {2, 3}:
+            raise ValueError('order should be 2 or 3')
+        forward = t_end > t_start
+        if not forward and eta:
+            raise ValueError('eta must be 0 for reverse sampling')
+        h_init = abs(h_init) * (1 if forward else -1)
+        atol = torch.tensor(atol)
+        rtol = torch.tensor(rtol)
+        s = t_start
+        x_prev = x
+        accept = True
+        pid = PIDStepSizeController(h_init, pcoeff, icoeff, dcoeff, 1.5 if eta else order, accept_safety)
+        info = {'steps': 0, 'nfe': 0, 'n_accept': 0, 'n_reject': 0}
+
+        while s < t_end - 1e-5 if forward else s > t_end + 1e-5:
+            eps_cache = {}
+            t = torch.minimum(t_end, s + pid.h) if forward else torch.maximum(t_end, s + pid.h)
+            if eta:
+                sd, su = get_ancestral_step(self.sigma(s), self.sigma(t), eta)
+                t_ = torch.minimum(t_end, self.t(sd))
+                su = (self.sigma(t) ** 2 - self.sigma(t_) ** 2) ** 0.5
+            else:
+                t_, su = t, 0.
+
+            eps, eps_cache = self.eps(eps_cache, 'eps', x, s)
+            denoised = x - self.sigma(s) * eps
+
+            if order == 2:
+                x_low, eps_cache = self.dpm_solver_1_step(x, s, t_, eps_cache=eps_cache)
+                x_high, eps_cache = self.dpm_solver_2_step(x, s, t_, eps_cache=eps_cache)
+            else:
+                x_low, eps_cache = self.dpm_solver_2_step(x, s, t_, r1=1 / 3, eps_cache=eps_cache)
+                x_high, eps_cache = self.dpm_solver_3_step(x, s, t_, eps_cache=eps_cache)
+            delta = torch.maximum(atol, rtol * torch.maximum(x_low.abs(), x_prev.abs()))
+            error = torch.linalg.norm((x_low - x_high) / delta) / x.numel() ** 0.5
+            accept = pid.propose_step(error)
+            if accept:
+                x_prev = x_low
+                x = x_high + su * s_noise * noise_sampler(self.sigma(s), self.sigma(t))
+                s = t
+                info['n_accept'] += 1
+            else:
+                info['n_reject'] += 1
+            info['nfe'] += order
+            info['steps'] += 1
+
+            if self.info_callback is not None:
+                self.info_callback({'x': x, 'i': info['steps'] - 1, 't': s, 't_up': s, 'denoised': denoised, 'error': error, 'h': pid.h, **info})
+
+        return x, info
+
+
+@torch.no_grad()
+def sample_dpm_fast(model, x, sigma_min, sigma_max, n, extra_args=None, callback=None, disable=None, eta=0., s_noise=1., noise_sampler=None):
+    """DPM-Solver-Fast (fixed step size). See https://arxiv.org/abs/2206.00927."""
+    if sigma_min <= 0 or sigma_max <= 0:
+        raise ValueError('sigma_min and sigma_max must not be 0')
+    with tqdm(total=n, disable=disable) as pbar:
+        dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update)
+        if callback is not None:
+            dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info})
+        return dpm_solver.dpm_solver_fast(x, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), n, eta, s_noise, noise_sampler)
+
+
+@torch.no_grad()
+def sample_dpm_adaptive(model, x, sigma_min, sigma_max, extra_args=None, callback=None, disable=None, order=3, rtol=0.05, atol=0.0078, h_init=0.05, pcoeff=0., icoeff=1., dcoeff=0., accept_safety=0.81, eta=0., s_noise=1., noise_sampler=None, return_info=False):
+    """DPM-Solver-12 and 23 (adaptive step size). See https://arxiv.org/abs/2206.00927."""
+    if sigma_min <= 0 or sigma_max <= 0:
+        raise ValueError('sigma_min and sigma_max must not be 0')
+    with tqdm(disable=disable) as pbar:
+        dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update)
+        if callback is not None:
+            dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info})
+        x, info = dpm_solver.dpm_solver_adaptive(x, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise, noise_sampler)
+    if return_info:
+        return x, info
+    return x
+
+
+@torch.no_grad()
+def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    """Ancestral sampling with DPM-Solver++(2S) second-order steps."""
+    extra_args = {} if extra_args is None else extra_args
+    noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+    sigma_fn = lambda t: t.neg().exp()
+    t_fn = lambda sigma: sigma.log().neg()
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigma_down == 0:
+            # Euler method
+            d = to_d(x, sigmas[i], denoised)
+            dt = sigma_down - sigmas[i]
+            x = x + d * dt
+        else:
+            # DPM-Solver++(2S)
+            t, t_next = t_fn(sigmas[i]), t_fn(sigma_down)
+            r = 1 / 2
+            h = t_next - t
+            s = t + r * h
+            x_2 = (sigma_fn(s) / sigma_fn(t)) * x - (-h * r).expm1() * denoised
+            denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
+            x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_2
+        # Noise addition
+        if sigmas[i + 1] > 0:
+            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
+    return x
+
+
+@torch.no_grad()
+def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
+    """DPM-Solver++ (stochastic)."""
+    sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+    noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max) if noise_sampler is None else noise_sampler
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    sigma_fn = lambda t: t.neg().exp()
+    t_fn = lambda sigma: sigma.log().neg()
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            # Euler method
+            d = to_d(x, sigmas[i], denoised)
+            dt = sigmas[i + 1] - sigmas[i]
+            x = x + d * dt
+        else:
+            # DPM-Solver++
+            t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
+            h = t_next - t
+            s = t + h * r
+            fac = 1 / (2 * r)
+
+            # Step 1
+            sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(s), eta)
+            s_ = t_fn(sd)
+            x_2 = (sigma_fn(s_) / sigma_fn(t)) * x - (t - s_).expm1() * denoised
+            x_2 = x_2 + noise_sampler(sigma_fn(t), sigma_fn(s)) * s_noise * su
+            denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
+
+            # Step 2
+            sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(t_next), eta)
+            t_next_ = t_fn(sd)
+            denoised_d = (1 - fac) * denoised + fac * denoised_2
+            x = (sigma_fn(t_next_) / sigma_fn(t)) * x - (t - t_next_).expm1() * denoised_d
+            x = x + noise_sampler(sigma_fn(t), sigma_fn(t_next)) * s_noise * su
+    return x
+
+
+@torch.no_grad()
+def sample_dpmpp_2m(model, x, sigmas, extra_args=None, callback=None, disable=None):
+    """DPM-Solver++(2M)."""
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    sigma_fn = lambda t: t.neg().exp()
+    t_fn = lambda sigma: sigma.log().neg()
+    old_denoised = None
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
+        h = t_next - t
+        if old_denoised is None or sigmas[i + 1] == 0:
+            x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised
+        else:
+            h_last = t - t_fn(sigmas[i - 1])
+            r = h_last / h
+            denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised
+            x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_d
+        old_denoised = denoised
+    return x
+
+
+@torch.no_grad()
+def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
+    """DPM-Solver++(2M) SDE."""
+
+    if solver_type not in {'heun', 'midpoint'}:
+        raise ValueError('solver_type must be \'heun\' or \'midpoint\'')
+
+    sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+    noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max) if noise_sampler is None else noise_sampler
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+
+    old_denoised = None
+    h_last = None
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            # DPM-Solver++(2M) SDE
+            t, s = -sigmas[i].log(), -sigmas[i + 1].log()
+            h = s - t
+            eta_h = eta * h
+
+            x = sigmas[i + 1] / sigmas[i] * (-eta_h).exp() * x + (-h - eta_h).expm1().neg() * denoised
+
+            if old_denoised is not None:
+                r = h_last / h
+                if solver_type == 'heun':
+                    x = x + ((-h - eta_h).expm1().neg() / (-h - eta_h) + 1) * (1 / r) * (denoised - old_denoised)
+                elif solver_type == 'midpoint':
+                    x = x + 0.5 * (-h - eta_h).expm1().neg() * (1 / r) * (denoised - old_denoised)
+
+            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * eta_h).expm1().neg().sqrt() * s_noise
+
+        old_denoised = denoised
+        h_last = h
+    return x
+
+
+@torch.no_grad()
+def sample_dpmpp_2m_test(model, x, sigmas, extra_args=None, callback=None, disable=None):
+    """DPM-Solver++(2M)."""
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    sigma_fn = lambda t: t.neg().exp()
+    t_fn = lambda sigma: sigma.log().neg()
+    old_denoised = None
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
+        h = t_next - t
+        
+        t_min = min(sigma_fn(t_next), sigma_fn(t))
+        t_max = max(sigma_fn(t_next), sigma_fn(t))
+
+        if old_denoised is None or sigmas[i + 1] == 0:
+            x = (t_min / t_max) * x - (-h).expm1() * denoised
+        else:
+            h_last = t - t_fn(sigmas[i - 1])
+
+            h_min = min(h_last, h)
+            h_max = max(h_last, h)
+            r = h_max / h_min
+
+            h_d = (h_max + h_min) / 2
+            denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised
+            x = (t_min / t_max) * x - (-h_d).expm1() * denoised_d
+
+        old_denoised = denoised
+    return x

dpm2mv2/sd_samplers_kdiffusion.py

@@ -0,0 +1,394 @@
+from collections import deque
+import torch
+import inspect
+import k_diffusion.sampling
+from modules import prompt_parser, devices, sd_samplers_common
+
+from modules.shared import opts, state
+import modules.shared as shared
+from modules.script_callbacks import CFGDenoiserParams, cfg_denoiser_callback
+from modules.script_callbacks import CFGDenoisedParams, cfg_denoised_callback
+from modules.script_callbacks import AfterCFGCallbackParams, cfg_after_cfg_callback
+
+samplers_k_diffusion = [
+    ('Euler a', 'sample_euler_ancestral', ['k_euler_a', 'k_euler_ancestral'], {"uses_ensd": True}),
+    ('Euler', 'sample_euler', ['k_euler'], {}),
+    ('LMS', 'sample_lms', ['k_lms'], {}),
+    ('Heun', 'sample_heun', ['k_heun'], {"second_order": True}),
+    ('DPM2', 'sample_dpm_2', ['k_dpm_2'], {'discard_next_to_last_sigma': True}),
+    ('DPM2 a', 'sample_dpm_2_ancestral', ['k_dpm_2_a'], {'discard_next_to_last_sigma': True, "uses_ensd": True}),
+    ('DPM++ 2S a', 'sample_dpmpp_2s_ancestral', ['k_dpmpp_2s_a'], {"uses_ensd": True, "second_order": True}),
+    ('DPM++ 2M', 'sample_dpmpp_2m', ['k_dpmpp_2m'], {}),
+    ('DPM++ 2M V2', 'sample_dpmpp_2m_test', ['k_dpmpp_2m'], {}),
+    ('DPM++ SDE', 'sample_dpmpp_sde', ['k_dpmpp_sde'], {"second_order": True, "brownian_noise": True}),
+    ('DPM++ 2M SDE', 'sample_dpmpp_2m_sde', ['k_dpmpp_2m_sde_ka'], {"brownian_noise": True, 'discard_next_to_last_sigma': True}),
+    ('DPM fast', 'sample_dpm_fast', ['k_dpm_fast'], {"uses_ensd": True}),
+    ('DPM adaptive', 'sample_dpm_adaptive', ['k_dpm_ad'], {"uses_ensd": True}),
+    ('LMS Karras', 'sample_lms', ['k_lms_ka'], {'scheduler': 'karras'}),
+    ('DPM2 Karras', 'sample_dpm_2', ['k_dpm_2_ka'], {'scheduler': 'karras', 'discard_next_to_last_sigma': True, "uses_ensd": True, "second_order": True}),
+    ('DPM2 a Karras', 'sample_dpm_2_ancestral', ['k_dpm_2_a_ka'], {'scheduler': 'karras', 'discard_next_to_last_sigma': True, "uses_ensd": True, "second_order": True}),
+    ('DPM++ 2S a Karras', 'sample_dpmpp_2s_ancestral', ['k_dpmpp_2s_a_ka'], {'scheduler': 'karras', "uses_ensd": True, "second_order": True}),
+    ('DPM++ 2M Karras', 'sample_dpmpp_2m', ['k_dpmpp_2m_ka'], {'scheduler': 'karras'}),
+    ('DPM++ 2M Karras V2', 'sample_dpmpp_2m_test', ['k_dpmpp_2m_ka'], {'scheduler': 'karras'}),
+    ('DPM++ SDE Karras', 'sample_dpmpp_sde', ['k_dpmpp_sde_ka'], {'scheduler': 'karras', "second_order": True, "brownian_noise": True}),
+    ('DPM++ 2M SDE Karras', 'sample_dpmpp_2m_sde', ['k_dpmpp_2m_sde_ka'], {'scheduler': 'karras', "brownian_noise": True, 'discard_next_to_last_sigma': True}),
+]
+
+samplers_data_k_diffusion = [
+    sd_samplers_common.SamplerData(label, lambda model, funcname=funcname: KDiffusionSampler(funcname, model), aliases, options)
+    for label, funcname, aliases, options in samplers_k_diffusion
+    if hasattr(k_diffusion.sampling, funcname)
+]
+
+sampler_extra_params = {
+    'sample_euler': ['s_churn', 's_tmin', 's_tmax', 's_noise'],
+    'sample_heun': ['s_churn', 's_tmin', 's_tmax', 's_noise'],
+    'sample_dpm_2': ['s_churn', 's_tmin', 's_tmax', 's_noise'],
+}
+
+
+class CFGDenoiser(torch.nn.Module):
+    """
+    Classifier free guidance denoiser. A wrapper for stable diffusion model (specifically for unet)
+    that can take a noisy picture and produce a noise-free picture using two guidances (prompts)
+    instead of one. Originally, the second prompt is just an empty string, but we use non-empty
+    negative prompt.
+    """
+
+    def __init__(self, model):
+        super().__init__()
+        self.inner_model = model
+        self.mask = None
+        self.nmask = None
+        self.init_latent = None
+        self.step = 0
+        self.image_cfg_scale = None
+
+    def combine_denoised(self, x_out, conds_list, uncond, cond_scale):
+        denoised_uncond = x_out[-uncond.shape[0]:]
+        denoised = torch.clone(denoised_uncond)
+
+        for i, conds in enumerate(conds_list):
+            for cond_index, weight in conds:
+                denoised[i] += (x_out[cond_index] - denoised_uncond[i]) * (weight * cond_scale)
+
+        return denoised
+
+    def combine_denoised_for_edit_model(self, x_out, cond_scale):
+        out_cond, out_img_cond, out_uncond = x_out.chunk(3)
+        denoised = out_uncond + cond_scale * (out_cond - out_img_cond) + self.image_cfg_scale * (out_img_cond - out_uncond)
+
+        return denoised
+
+    def forward(self, x, sigma, uncond, cond, cond_scale, s_min_uncond, image_cond):
+        if state.interrupted or state.skipped:
+            raise sd_samplers_common.InterruptedException
+
+        # at self.image_cfg_scale == 1.0 produced results for edit model are the same as with normal sampling,
+        # so is_edit_model is set to False to support AND composition.
+        is_edit_model = shared.sd_model.cond_stage_key == "edit" and self.image_cfg_scale is not None and self.image_cfg_scale != 1.0
+
+        conds_list, tensor = prompt_parser.reconstruct_multicond_batch(cond, self.step)
+        uncond = prompt_parser.reconstruct_cond_batch(uncond, self.step)
+
+        assert not is_edit_model or all(len(conds) == 1 for conds in conds_list), "AND is not supported for InstructPix2Pix checkpoint (unless using Image CFG scale = 1.0)"
+
+        batch_size = len(conds_list)
+        repeats = [len(conds_list[i]) for i in range(batch_size)]
+
+        if shared.sd_model.model.conditioning_key == "crossattn-adm":
+            image_uncond = torch.zeros_like(image_cond)
+            make_condition_dict = lambda c_crossattn, c_adm: {"c_crossattn": c_crossattn, "c_adm": c_adm}
+        else:
+            image_uncond = image_cond
+            make_condition_dict = lambda c_crossattn, c_concat: {"c_crossattn": c_crossattn, "c_concat": [c_concat]}
+
+        if not is_edit_model:
+            x_in = torch.cat([torch.stack([x[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [x])
+            sigma_in = torch.cat([torch.stack([sigma[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [sigma])
+            image_cond_in = torch.cat([torch.stack([image_cond[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [image_uncond])
+        else:
+            x_in = torch.cat([torch.stack([x[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [x] + [x])
+            sigma_in = torch.cat([torch.stack([sigma[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [sigma] + [sigma])
+            image_cond_in = torch.cat([torch.stack([image_cond[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [image_uncond] + [torch.zeros_like(self.init_latent)])
+
+        denoiser_params = CFGDenoiserParams(x_in, image_cond_in, sigma_in, state.sampling_step, state.sampling_steps, tensor, uncond)
+        cfg_denoiser_callback(denoiser_params)
+        x_in = denoiser_params.x
+        image_cond_in = denoiser_params.image_cond
+        sigma_in = denoiser_params.sigma
+        tensor = denoiser_params.text_cond
+        uncond = denoiser_params.text_uncond
+        skip_uncond = False
+
+        # alternating uncond allows for higher thresholds without the quality loss normally expected from raising it
+        if self.step % 2 and s_min_uncond > 0 and sigma[0] < s_min_uncond and not is_edit_model:
+            skip_uncond = True
+            x_in = x_in[:-batch_size]
+            sigma_in = sigma_in[:-batch_size]
+
+        if tensor.shape[1] == uncond.shape[1] or skip_uncond:
+            if is_edit_model:
+                cond_in = torch.cat([tensor, uncond, uncond])
+            elif skip_uncond:
+                cond_in = tensor
+            else:
+                cond_in = torch.cat([tensor, uncond])
+
+            if shared.batch_cond_uncond:
+                x_out = self.inner_model(x_in, sigma_in, cond=make_condition_dict([cond_in], image_cond_in))
+            else:
+                x_out = torch.zeros_like(x_in)
+                for batch_offset in range(0, x_out.shape[0], batch_size):
+                    a = batch_offset
+                    b = a + batch_size
+                    x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=make_condition_dict([cond_in[a:b]], image_cond_in[a:b]))
+        else:
+            x_out = torch.zeros_like(x_in)
+            batch_size = batch_size*2 if shared.batch_cond_uncond else batch_size
+            for batch_offset in range(0, tensor.shape[0], batch_size):
+                a = batch_offset
+                b = min(a + batch_size, tensor.shape[0])
+
+                if not is_edit_model:
+                    c_crossattn = [tensor[a:b]]
+                else:
+                    c_crossattn = torch.cat([tensor[a:b]], uncond)
+
+                x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=make_condition_dict(c_crossattn, image_cond_in[a:b]))
+
+            if not skip_uncond:
+                x_out[-uncond.shape[0]:] = self.inner_model(x_in[-uncond.shape[0]:], sigma_in[-uncond.shape[0]:], cond=make_condition_dict([uncond], image_cond_in[-uncond.shape[0]:]))
+
+        denoised_image_indexes = [x[0][0] for x in conds_list]
+        if skip_uncond:
+            fake_uncond = torch.cat([x_out[i:i+1] for i in denoised_image_indexes])
+            x_out = torch.cat([x_out, fake_uncond])  # we skipped uncond denoising, so we put cond-denoised image to where the uncond-denoised image should be
+
+        denoised_params = CFGDenoisedParams(x_out, state.sampling_step, state.sampling_steps, self.inner_model)
+        cfg_denoised_callback(denoised_params)
+
+        devices.test_for_nans(x_out, "unet")
+
+        if opts.live_preview_content == "Prompt":
+            sd_samplers_common.store_latent(torch.cat([x_out[i:i+1] for i in denoised_image_indexes]))
+        elif opts.live_preview_content == "Negative prompt":
+            sd_samplers_common.store_latent(x_out[-uncond.shape[0]:])
+
+        if is_edit_model:
+            denoised = self.combine_denoised_for_edit_model(x_out, cond_scale)
+        elif skip_uncond:
+            denoised = self.combine_denoised(x_out, conds_list, uncond, 1.0)
+        else:
+            denoised = self.combine_denoised(x_out, conds_list, uncond, cond_scale)
+
+        if self.mask is not None:
+            denoised = self.init_latent * self.mask + self.nmask * denoised
+
+        after_cfg_callback_params = AfterCFGCallbackParams(denoised, state.sampling_step, state.sampling_steps)
+        cfg_after_cfg_callback(after_cfg_callback_params)
+        denoised = after_cfg_callback_params.x
+
+        self.step += 1
+        return denoised
+
+
+class TorchHijack:
+    def __init__(self, sampler_noises):
+        # Using a deque to efficiently receive the sampler_noises in the same order as the previous index-based
+        # implementation.
+        self.sampler_noises = deque(sampler_noises)
+
+    def __getattr__(self, item):
+        if item == 'randn_like':
+            return self.randn_like
+
+        if hasattr(torch, item):
+            return getattr(torch, item)
+
+        raise AttributeError(f"'{type(self).__name__}' object has no attribute '{item}'")
+
+    def randn_like(self, x):
+        if self.sampler_noises:
+            noise = self.sampler_noises.popleft()
+            if noise.shape == x.shape:
+                return noise
+
+        if opts.randn_source == "CPU" or x.device.type == 'mps':
+            return torch.randn_like(x, device=devices.cpu).to(x.device)
+        else:
+            return torch.randn_like(x)
+
+
+class KDiffusionSampler:
+    def __init__(self, funcname, sd_model):
+        denoiser = k_diffusion.external.CompVisVDenoiser if sd_model.parameterization == "v" else k_diffusion.external.CompVisDenoiser
+
+        self.model_wrap = denoiser(sd_model, quantize=shared.opts.enable_quantization)
+        self.funcname = funcname
+        self.func = getattr(k_diffusion.sampling, self.funcname)
+        self.extra_params = sampler_extra_params.get(funcname, [])
+        self.model_wrap_cfg = CFGDenoiser(self.model_wrap)
+        self.sampler_noises = None
+        self.stop_at = None
+        self.eta = None
+        self.config = None  # set by the function calling the constructor
+        self.last_latent = None
+        self.s_min_uncond = None
+
+        self.conditioning_key = sd_model.model.conditioning_key
+
+    def callback_state(self, d):
+        step = d['i']
+        latent = d["denoised"]
+        if opts.live_preview_content == "Combined":
+            sd_samplers_common.store_latent(latent)
+        self.last_latent = latent
+
+        if self.stop_at is not None and step > self.stop_at:
+            raise sd_samplers_common.InterruptedException
+
+        state.sampling_step = step
+        shared.total_tqdm.update()
+
+    def launch_sampling(self, steps, func):
+        state.sampling_steps = steps
+        state.sampling_step = 0
+
+        try:
+            return func()
+        except sd_samplers_common.InterruptedException:
+            return self.last_latent
+
+    def number_of_needed_noises(self, p):
+        return p.steps
+
+    def initialize(self, p):
+        self.model_wrap_cfg.mask = p.mask if hasattr(p, 'mask') else None
+        self.model_wrap_cfg.nmask = p.nmask if hasattr(p, 'nmask') else None
+        self.model_wrap_cfg.step = 0
+        self.model_wrap_cfg.image_cfg_scale = getattr(p, 'image_cfg_scale', None)
+        self.eta = p.eta if p.eta is not None else opts.eta_ancestral
+        self.s_min_uncond = getattr(p, 's_min_uncond', 0.0)
+
+        k_diffusion.sampling.torch = TorchHijack(self.sampler_noises if self.sampler_noises is not None else [])
+
+        extra_params_kwargs = {}
+        for param_name in self.extra_params:
+            if hasattr(p, param_name) and param_name in inspect.signature(self.func).parameters:
+                extra_params_kwargs[param_name] = getattr(p, param_name)
+
+        if 'eta' in inspect.signature(self.func).parameters:
+            if self.eta != 1.0:
+                p.extra_generation_params["Eta"] = self.eta
+
+            extra_params_kwargs['eta'] = self.eta
+
+        return extra_params_kwargs
+
+    def get_sigmas(self, p, steps):
+        discard_next_to_last_sigma = self.config is not None and self.config.options.get('discard_next_to_last_sigma', False)
+        if opts.always_discard_next_to_last_sigma and not discard_next_to_last_sigma:
+            discard_next_to_last_sigma = True
+            p.extra_generation_params["Discard penultimate sigma"] = True
+
+        steps += 1 if discard_next_to_last_sigma else 0
+
+        if p.sampler_noise_scheduler_override:
+            sigmas = p.sampler_noise_scheduler_override(steps)
+        elif self.config is not None and self.config.options.get('scheduler', None) == 'karras':
+            sigma_min, sigma_max = (0.1, 10) if opts.use_old_karras_scheduler_sigmas else (self.model_wrap.sigmas[0].item(), self.model_wrap.sigmas[-1].item())
+
+            sigmas = k_diffusion.sampling.get_sigmas_karras(n=steps, sigma_min=sigma_min, sigma_max=sigma_max, device=shared.device)
+        else:
+            sigmas = self.model_wrap.get_sigmas(steps)
+
+        if discard_next_to_last_sigma:
+            sigmas = torch.cat([sigmas[:-2], sigmas[-1:]])
+
+        return sigmas
+
+    def create_noise_sampler(self, x, sigmas, p):
+        """For DPM++ SDE: manually create noise sampler to enable deterministic results across different batch sizes"""
+        if shared.opts.no_dpmpp_sde_batch_determinism:
+            return None
+
+        from k_diffusion.sampling import BrownianTreeNoiseSampler
+        sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+        current_iter_seeds = p.all_seeds[p.iteration * p.batch_size:(p.iteration + 1) * p.batch_size]
+        return BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=current_iter_seeds)
+
+    def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None, image_conditioning=None):
+        steps, t_enc = sd_samplers_common.setup_img2img_steps(p, steps)
+
+        sigmas = self.get_sigmas(p, steps)
+
+        sigma_sched = sigmas[steps - t_enc - 1:]
+        xi = x + noise * sigma_sched[0]
+
+        extra_params_kwargs = self.initialize(p)
+        parameters = inspect.signature(self.func).parameters
+
+        if 'sigma_min' in parameters:
+            ## last sigma is zero which isn't allowed by DPM Fast & Adaptive so taking value before last
+            extra_params_kwargs['sigma_min'] = sigma_sched[-2]
+        if 'sigma_max' in parameters:
+            extra_params_kwargs['sigma_max'] = sigma_sched[0]
+        if 'n' in parameters:
+            extra_params_kwargs['n'] = len(sigma_sched) - 1
+        if 'sigma_sched' in parameters:
+            extra_params_kwargs['sigma_sched'] = sigma_sched
+        if 'sigmas' in parameters:
+            extra_params_kwargs['sigmas'] = sigma_sched
+
+        if self.config.options.get('brownian_noise', False):
+            noise_sampler = self.create_noise_sampler(x, sigmas, p)
+            extra_params_kwargs['noise_sampler'] = noise_sampler
+
+        self.model_wrap_cfg.init_latent = x
+        self.last_latent = x
+        extra_args = {
+            'cond': conditioning,
+            'image_cond': image_conditioning,
+            'uncond': unconditional_conditioning,
+            'cond_scale': p.cfg_scale,
+            's_min_uncond': self.s_min_uncond
+        }
+
+        samples = self.launch_sampling(t_enc + 1, lambda: self.func(self.model_wrap_cfg, xi, extra_args=extra_args, disable=False, callback=self.callback_state, **extra_params_kwargs))
+
+        return samples
+
+    def sample(self, p, x, conditioning, unconditional_conditioning, steps=None, image_conditioning=None):
+        steps = steps or p.steps
+
+        sigmas = self.get_sigmas(p, steps)
+
+        x = x * sigmas[0]
+
+        extra_params_kwargs = self.initialize(p)
+        parameters = inspect.signature(self.func).parameters
+
+        if 'sigma_min' in parameters:
+            extra_params_kwargs['sigma_min'] = self.model_wrap.sigmas[0].item()
+            extra_params_kwargs['sigma_max'] = self.model_wrap.sigmas[-1].item()
+            if 'n' in parameters:
+                extra_params_kwargs['n'] = steps
+        else:
+            extra_params_kwargs['sigmas'] = sigmas
+
+        if self.config.options.get('brownian_noise', False):
+            noise_sampler = self.create_noise_sampler(x, sigmas, p)
+            extra_params_kwargs['noise_sampler'] = noise_sampler
+
+        self.last_latent = x
+        samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, x, extra_args={
+            'cond': conditioning,
+            'image_cond': image_conditioning,
+            'uncond': unconditional_conditioning,
+            'cond_scale': p.cfg_scale,
+            's_min_uncond': self.s_min_uncond
+        }, disable=False, callback=self.callback_state, **extra_params_kwargs))
+
+        return samples
+