mirror of https://github.com/vladmandic/automatic
451 lines
18 KiB
Python
451 lines
18 KiB
Python
# Copyright 2025 The RES4LYF Team (Clybius) and The HuggingFace Team. All rights reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
from typing import ClassVar, Literal
|
|
|
|
import numpy as np
|
|
import torch
|
|
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
|
from diffusers.schedulers.scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
|
|
|
|
from diffusers.utils import logging
|
|
|
|
from .phi_functions import Phi, calculate_gamma
|
|
|
|
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
|
|
|
|
|
class RESMultistepScheduler(SchedulerMixin, ConfigMixin):
|
|
"""
|
|
RESMultistepScheduler (Restartable Exponential Integrator) ported from RES4LYF.
|
|
|
|
Supports RES 2M, 3M and DEIS 2M, 3M variants.
|
|
|
|
Args:
|
|
num_train_timesteps (`int`, defaults to 1000):
|
|
The number of diffusion steps to train the model.
|
|
beta_start (`float`, defaults to 0.0001):
|
|
The starting `beta` value of inference.
|
|
beta_end (`float`, defaults to 0.02):
|
|
The final `beta` value.
|
|
beta_schedule (`str`, defaults to "linear"):
|
|
The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model.
|
|
prediction_type (`str`, defaults to "epsilon"):
|
|
The prediction type of the scheduler function.
|
|
variant (`str`, defaults to "res_2m"):
|
|
The specific RES/DEIS variant to use. Supported: "res_2m", "res_3m", "deis_2m", "deis_3m".
|
|
use_analytic_solution (`bool`, defaults to True):
|
|
Whether to use high-precision analytic solutions for phi functions.
|
|
"""
|
|
|
|
_compatibles: ClassVar[list[str]] = [e.name for e in KarrasDiffusionSchedulers]
|
|
order = 1
|
|
|
|
@register_to_config
|
|
def __init__(
|
|
self,
|
|
num_train_timesteps: int = 1000,
|
|
beta_start: float = 0.00085,
|
|
beta_end: float = 0.012,
|
|
beta_schedule: str = "linear",
|
|
prediction_type: str = "epsilon",
|
|
variant: Literal["res_2m", "res_3m", "deis_2m", "deis_3m"] = "res_2m",
|
|
use_analytic_solution: bool = True,
|
|
timestep_spacing: str = "linspace",
|
|
steps_offset: int = 0,
|
|
rescale_betas_zero_snr: bool = False,
|
|
use_karras_sigmas: bool = False,
|
|
use_exponential_sigmas: bool = False,
|
|
use_beta_sigmas: bool = False,
|
|
use_flow_sigmas: bool = False,
|
|
shift: float = 1.0,
|
|
use_dynamic_shifting: bool = False,
|
|
base_shift: float = 0.5,
|
|
max_shift: float = 1.15,
|
|
base_image_seq_len: int = 256,
|
|
max_image_seq_len: int = 4096,
|
|
):
|
|
from .scheduler_utils import betas_for_alpha_bar, rescale_zero_terminal_snr
|
|
|
|
if beta_schedule == "linear":
|
|
self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
|
|
elif beta_schedule == "scaled_linear":
|
|
self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
|
|
elif beta_schedule == "squaredcos_cap_v2":
|
|
self.betas = betas_for_alpha_bar(num_train_timesteps)
|
|
else:
|
|
raise NotImplementedError(f"{beta_schedule} is not implemented")
|
|
|
|
if rescale_betas_zero_snr:
|
|
self.betas = rescale_zero_terminal_snr(self.betas)
|
|
|
|
self.alphas = 1.0 - self.betas
|
|
self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
|
|
|
|
# Buffer for multistep
|
|
self.model_outputs = []
|
|
self.x0_outputs = []
|
|
self.prev_sigmas = []
|
|
self._step_index = None
|
|
self._begin_index = None
|
|
self.init_noise_sigma = 1.0
|
|
|
|
@property
|
|
def step_index(self) -> int | None:
|
|
return self._step_index
|
|
|
|
@property
|
|
def begin_index(self) -> int | None:
|
|
return self._begin_index
|
|
|
|
def set_begin_index(self, begin_index: int = 0) -> None:
|
|
self._begin_index = begin_index
|
|
|
|
def scale_model_input(self, sample: torch.Tensor, timestep: float | torch.Tensor) -> torch.Tensor:
|
|
if self._step_index is None:
|
|
self._init_step_index(timestep)
|
|
if self.config.prediction_type == "flow_prediction":
|
|
return sample
|
|
sigma = self.sigmas[self._step_index]
|
|
return sample / ((sigma**2 + 1) ** 0.5)
|
|
|
|
def set_timesteps(self, num_inference_steps: int, device: str | torch.device = None, mu: float | None = None, dtype: torch.dtype = torch.float32):
|
|
from .scheduler_utils import (
|
|
apply_shift,
|
|
get_dynamic_shift,
|
|
get_sigmas_beta,
|
|
get_sigmas_exponential,
|
|
get_sigmas_karras,
|
|
)
|
|
|
|
self.num_inference_steps = num_inference_steps
|
|
|
|
if self.config.timestep_spacing == "linspace":
|
|
timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=float)[::-1].copy()
|
|
elif self.config.timestep_spacing == "leading":
|
|
step_ratio = self.config.num_train_timesteps // self.num_inference_steps
|
|
timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy()
|
|
timesteps += self.config.steps_offset
|
|
elif self.config.timestep_spacing == "trailing":
|
|
step_ratio = self.config.num_train_timesteps / self.num_inference_steps
|
|
timesteps = (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy()
|
|
timesteps -= 1
|
|
else:
|
|
raise ValueError(f"timestep_spacing {self.config.timestep_spacing} is not supported.")
|
|
|
|
sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
|
|
# Linear remapping for Flow Matching
|
|
if self.config.use_flow_sigmas:
|
|
# Standardize linear spacing
|
|
sigmas = np.linspace(1.0, 1 / 1000, num_inference_steps)
|
|
else:
|
|
sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
|
|
|
|
if self.config.use_karras_sigmas:
|
|
sigmas = get_sigmas_karras(num_inference_steps, sigmas[-1], sigmas[0], device=device, dtype=dtype).cpu().numpy()
|
|
elif self.config.use_exponential_sigmas:
|
|
sigmas = get_sigmas_exponential(num_inference_steps, sigmas[-1], sigmas[0], device=device, dtype=dtype).cpu().numpy()
|
|
elif self.config.use_beta_sigmas:
|
|
sigmas = get_sigmas_beta(num_inference_steps, sigmas[-1], sigmas[0], device=device, dtype=dtype).cpu().numpy()
|
|
elif self.config.use_flow_sigmas:
|
|
# Already handled above, ensuring variable consistency
|
|
sigmas = np.linspace(1.0, 1 / 1000, num_inference_steps)
|
|
|
|
if self.config.shift != 1.0 or self.config.use_dynamic_shifting:
|
|
shift = self.config.shift
|
|
if self.config.use_dynamic_shifting and mu is not None:
|
|
shift = get_dynamic_shift(
|
|
mu,
|
|
self.config.base_shift,
|
|
self.config.max_shift,
|
|
self.config.base_image_seq_len,
|
|
self.config.max_image_seq_len,
|
|
)
|
|
sigmas = apply_shift(torch.from_numpy(sigmas), shift).numpy()
|
|
|
|
if self.config.use_flow_sigmas:
|
|
timesteps = sigmas * self.config.num_train_timesteps
|
|
|
|
self.sigmas = torch.from_numpy(np.concatenate([sigmas, [0.0]])).to(device=device, dtype=dtype)
|
|
self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=dtype)
|
|
self.init_noise_sigma = self.sigmas.max().item() if self.sigmas.numel() > 0 else 1.0
|
|
|
|
self._step_index = None
|
|
self._begin_index = None
|
|
self.model_outputs = []
|
|
self.x0_outputs = []
|
|
self.prev_sigmas = []
|
|
|
|
self.lower_order_nums = 0
|
|
|
|
def index_for_timestep(self, timestep, schedule_timesteps=None):
|
|
from .scheduler_utils import index_for_timestep
|
|
if schedule_timesteps is None:
|
|
schedule_timesteps = self.timesteps
|
|
return index_for_timestep(timestep, schedule_timesteps)
|
|
|
|
def add_noise(
|
|
self,
|
|
original_samples: torch.Tensor,
|
|
noise: torch.Tensor,
|
|
timesteps: torch.Tensor,
|
|
) -> torch.Tensor:
|
|
from .scheduler_utils import add_noise_to_sample
|
|
return add_noise_to_sample(original_samples, noise, self.sigmas, timesteps, self.timesteps)
|
|
|
|
def step(
|
|
self,
|
|
model_output: torch.Tensor,
|
|
timestep: float | torch.Tensor,
|
|
sample: torch.Tensor,
|
|
return_dict: bool = True,
|
|
) -> SchedulerOutput | tuple:
|
|
if self._step_index is None:
|
|
self._init_step_index(timestep)
|
|
|
|
step = self._step_index
|
|
sigma = self.sigmas[step]
|
|
sigma_next = self.sigmas[step + 1]
|
|
|
|
h = -torch.log(sigma_next / sigma) if sigma > 0 and sigma_next > 0 else torch.zeros_like(sigma)
|
|
|
|
# RECONSTRUCT X0 (Matching PEC pattern)
|
|
if self.config.prediction_type == "epsilon":
|
|
x0 = sample - sigma * model_output
|
|
elif self.config.prediction_type == "sample":
|
|
x0 = model_output
|
|
elif self.config.prediction_type == "v_prediction":
|
|
alpha_t = 1.0 / (sigma**2 + 1) ** 0.5
|
|
sigma_t = sigma * alpha_t
|
|
x0 = alpha_t * sample - sigma_t * model_output
|
|
elif self.config.prediction_type == "flow_prediction":
|
|
x0 = sample - sigma * model_output
|
|
else:
|
|
raise ValueError(f"prediction_type {self.config.prediction_type} is not supported.")
|
|
|
|
self.model_outputs.append(model_output)
|
|
self.x0_outputs.append(x0)
|
|
self.prev_sigmas.append(sigma)
|
|
|
|
# Order logic
|
|
variant = self.config.variant
|
|
order = int(variant[-2]) if variant.endswith("m") else 1
|
|
|
|
# Effective order for current step
|
|
curr_order = min(len(self.prev_sigmas), order) if sigma > 0 else 1
|
|
|
|
if self.config.prediction_type == "flow_prediction":
|
|
# Variable Step Adams-Bashforth for Flow Matching
|
|
dt = sigma_next - sigma
|
|
v_n = model_output
|
|
|
|
if curr_order == 1:
|
|
x_next = sample + dt * v_n
|
|
elif curr_order == 2:
|
|
# AB2
|
|
sigma_prev = self.prev_sigmas[-2]
|
|
dt_prev = sigma - sigma_prev
|
|
r = dt / dt_prev if abs(dt_prev) > 1e-8 else 0.0
|
|
|
|
# Stability check
|
|
if dt_prev == 0 or r < -0.9 or r > 2.0: # Fallback
|
|
x_next = sample + dt * v_n
|
|
else:
|
|
c0 = 1 + 0.5 * r
|
|
c1 = -0.5 * r
|
|
x_next = sample + dt * (c0 * v_n + c1 * self.model_outputs[-2])
|
|
elif curr_order >= 3:
|
|
# Re-implement AB2 logic
|
|
sigma_prev = self.prev_sigmas[-2]
|
|
dt_prev = sigma - sigma_prev
|
|
r = dt / dt_prev if abs(dt_prev) > 1e-8 else 0.0
|
|
c0 = 1 + 0.5 * r
|
|
c1 = -0.5 * r
|
|
x_next = sample + dt * (c0 * v_n + c1 * self.model_outputs[-2])
|
|
|
|
self._step_index += 1
|
|
if len(self.model_outputs) > order:
|
|
self.model_outputs.pop(0)
|
|
self.x0_outputs.pop(0)
|
|
self.prev_sigmas.pop(0)
|
|
|
|
if not return_dict:
|
|
return (x_next,)
|
|
return SchedulerOutput(prev_sample=x_next)
|
|
|
|
# Exponential Integrator Setup
|
|
phi = Phi(h, [0], getattr(self.config, "use_analytic_solution", True))
|
|
phi_1 = phi(1)
|
|
|
|
if variant.startswith("res"):
|
|
# Force Order 1 at the end of schedule
|
|
if self.num_inference_steps is not None and self._step_index >= self.num_inference_steps - 3:
|
|
curr_order = 1
|
|
|
|
if curr_order == 2:
|
|
h_prev = -torch.log(self.prev_sigmas[-1] / self.prev_sigmas[-2])
|
|
elif curr_order == 3:
|
|
pass
|
|
else:
|
|
pass
|
|
|
|
# Exponential Integrator Update in x-space
|
|
if curr_order == 1:
|
|
res = phi_1 * x0
|
|
elif curr_order == 2:
|
|
# b2 = -phi_2 / r
|
|
# b2 = -phi_2 / r = -phi(2) / (h_prev/h)
|
|
# Here we use: b2 = phi(2) / ((-h_prev / h) + 1e-9)
|
|
# Since (-h_prev/h) is negative (-r), this gives correct negative sign for b2.
|
|
|
|
# Stability check
|
|
r_check = h_prev / (h + 1e-9) # This is effectively -r if using h_prev definition above?
|
|
# Wait, h_prev above is -log(). Positive.
|
|
# h is positive.
|
|
# So h_prev/h is positive. defined as r in other files.
|
|
# But here code uses -h_prev / h in denominator.
|
|
|
|
# Stability check
|
|
r_check = h_prev / (h + 1e-9)
|
|
|
|
# Hard Restart
|
|
if r_check < 0.5 or r_check > 2.0:
|
|
res = phi_1 * x0
|
|
else:
|
|
b2 = phi(2) / ((-h_prev / h) + 1e-9)
|
|
b1 = phi_1 - b2
|
|
res = b1 * self.x0_outputs[-1] + b2 * self.x0_outputs[-2]
|
|
elif curr_order == 3:
|
|
# Generalized AB3 for Exponential Integrators
|
|
h_p1 = -torch.log(self.prev_sigmas[-1] / (self.prev_sigmas[-2] + 1e-9))
|
|
h_p2 = -torch.log(self.prev_sigmas[-1] / (self.prev_sigmas[-3] + 1e-9))
|
|
r1 = h_p1 / (h + 1e-9)
|
|
r2 = h_p2 / (h + 1e-9)
|
|
|
|
if r1 < 0.5 or r1 > 2.0 or r2 < 0.5 or r2 > 2.0:
|
|
res = phi_1 * x0
|
|
else:
|
|
phi_2, phi_3 = phi(2), phi(3)
|
|
denom = r2 - r1 + 1e-9
|
|
b3 = (phi_3 + r1 * phi_2) / (r2 * denom)
|
|
b2 = -(phi_3 + r2 * phi_2) / (r1 * denom)
|
|
b1 = phi_1 - b2 - b3
|
|
res = b1 * self.x0_outputs[-1] + b2 * self.x0_outputs[-2] + b3 * self.x0_outputs[-3]
|
|
else:
|
|
res = phi_1 * x0
|
|
|
|
if sigma_next == 0:
|
|
x_next = x0
|
|
else:
|
|
x_next = torch.exp(-h) * sample + h * res
|
|
|
|
else:
|
|
# DEIS logic (Linear multistep in log-sigma space)
|
|
b = self._get_deis_coefficients(curr_order, sigma, sigma_next)
|
|
|
|
# For DEIS, we apply b to the denoised estimates
|
|
res = torch.zeros_like(sample)
|
|
for i, b_val in enumerate(b[0]):
|
|
idx = len(self.x0_outputs) - 1 - i
|
|
if idx >= 0:
|
|
res += b_val * self.x0_outputs[idx]
|
|
|
|
# DEIS update in x-space
|
|
if sigma_next == 0:
|
|
x_next = x0
|
|
else:
|
|
x_next = torch.exp(-h) * sample + h * res
|
|
|
|
self._step_index += 1
|
|
|
|
if len(self.model_outputs) > order:
|
|
self.model_outputs.pop(0)
|
|
self.x0_outputs.pop(0)
|
|
self.prev_sigmas.pop(0)
|
|
|
|
if not return_dict:
|
|
return (x_next,)
|
|
|
|
return SchedulerOutput(prev_sample=x_next)
|
|
|
|
def _get_res_coefficients(self, rk_type, h, c2, c3):
|
|
ci = [0, c2, c3]
|
|
phi = Phi(h, ci, getattr(self.config, "use_analytic_solution", True))
|
|
|
|
if rk_type == "res_2s":
|
|
b2 = phi(2) / (c2 + 1e-9)
|
|
b = [[phi(1) - b2, b2]]
|
|
a = [[0, 0], [c2 * phi(1, 2), 0]]
|
|
elif rk_type == "res_3s":
|
|
gamma_val = calculate_gamma(c2, c3)
|
|
b3 = phi(2) / (gamma_val * c2 + c3 + 1e-9)
|
|
b2 = gamma_val * b3
|
|
b = [[phi(1) - (b2 + b3), b2, b3]]
|
|
a = [] # Simplified
|
|
else:
|
|
b = [[phi(1)]]
|
|
a = [[0]]
|
|
return a, b, ci
|
|
|
|
def _get_deis_coefficients(self, order, sigma, sigma_next):
|
|
h = -torch.log(sigma_next / sigma) if sigma > 0 and sigma_next > 0 else torch.zeros_like(sigma)
|
|
phi = Phi(h, [0], getattr(self.config, "use_analytic_solution", True))
|
|
phi_1 = phi(1)
|
|
|
|
if order == 1:
|
|
return [[phi_1]]
|
|
elif order == 2:
|
|
h_prev = -torch.log(self.prev_sigmas[-1] / (self.prev_sigmas[-2] + 1e-9))
|
|
r = h_prev / (h + 1e-9)
|
|
|
|
# Correct Adams-Bashforth-like coefficients for Exponential Integrators
|
|
|
|
# Hard Restart for stability
|
|
if r < 0.5 or r > 2.0:
|
|
return [[phi_1]]
|
|
|
|
phi_2 = phi(2)
|
|
b2 = -phi_2 / (r + 1e-9)
|
|
b1 = phi_1 - b2
|
|
return [[b1, b2]]
|
|
elif order == 3:
|
|
h_prev1 = -torch.log(self.prev_sigmas[-1] / (self.prev_sigmas[-2] + 1e-9))
|
|
h_prev2 = -torch.log(self.prev_sigmas[-1] / (self.prev_sigmas[-3] + 1e-9))
|
|
r1 = h_prev1 / (h + 1e-9)
|
|
r2 = h_prev2 / (h + 1e-9)
|
|
|
|
if r1 < 0.5 or r1 > 2.0 or r2 < 0.5 or r2 > 2.0:
|
|
return [[phi_1]]
|
|
|
|
phi_2 = phi(2)
|
|
phi_3 = phi(3)
|
|
|
|
# Generalized AB3 for Exponential Integrators (Varying steps)
|
|
denom = r2 - r1 + 1e-9
|
|
b3 = (phi_3 + r1 * phi_2) / (r2 * denom)
|
|
b2 = -(phi_3 + r2 * phi_2) / (r1 * denom)
|
|
b1 = phi_1 - (b2 + b3)
|
|
return [[b1, b2, b3]]
|
|
else:
|
|
return [[phi_1]]
|
|
|
|
def _init_step_index(self, timestep):
|
|
if self.begin_index is None:
|
|
self._step_index = self.index_for_timestep(timestep)
|
|
else:
|
|
self._step_index = self._begin_index
|
|
|
|
def __len__(self):
|
|
return self.config.num_train_timesteps
|