sd_samplers_timesteps.py 6.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167
  1. import torch
  2. import inspect
  3. import sys
  4. from modules import devices, sd_samplers_common, sd_samplers_timesteps_impl
  5. from modules.sd_samplers_cfg_denoiser import CFGDenoiser
  6. from modules.script_callbacks import ExtraNoiseParams, extra_noise_callback
  7. from modules.shared import opts
  8. import modules.shared as shared
  9. samplers_timesteps = [
  10. ('DDIM', sd_samplers_timesteps_impl.ddim, ['ddim'], {}),
  11. ('DDIM CFG++', sd_samplers_timesteps_impl.ddim_cfgpp, ['ddim_cfgpp'], {}),
  12. ('PLMS', sd_samplers_timesteps_impl.plms, ['plms'], {}),
  13. ('UniPC', sd_samplers_timesteps_impl.unipc, ['unipc'], {}),
  14. ]
  15. samplers_data_timesteps = [
  16. sd_samplers_common.SamplerData(label, lambda model, funcname=funcname: CompVisSampler(funcname, model), aliases, options)
  17. for label, funcname, aliases, options in samplers_timesteps
  18. ]
  19. class CompVisTimestepsDenoiser(torch.nn.Module):
  20. def __init__(self, model, *args, **kwargs):
  21. super().__init__(*args, **kwargs)
  22. self.inner_model = model
  23. def forward(self, input, timesteps, **kwargs):
  24. return self.inner_model.apply_model(input, timesteps, **kwargs)
  25. class CompVisTimestepsVDenoiser(torch.nn.Module):
  26. def __init__(self, model, *args, **kwargs):
  27. super().__init__(*args, **kwargs)
  28. self.inner_model = model
  29. def predict_eps_from_z_and_v(self, x_t, t, v):
  30. return torch.sqrt(self.inner_model.alphas_cumprod)[t.to(torch.int), None, None, None] * v + torch.sqrt(1 - self.inner_model.alphas_cumprod)[t.to(torch.int), None, None, None] * x_t
  31. def forward(self, input, timesteps, **kwargs):
  32. model_output = self.inner_model.apply_model(input, timesteps, **kwargs)
  33. e_t = self.predict_eps_from_z_and_v(input, timesteps, model_output)
  34. return e_t
  35. class CFGDenoiserTimesteps(CFGDenoiser):
  36. def __init__(self, sampler):
  37. super().__init__(sampler)
  38. self.alphas = shared.sd_model.alphas_cumprod
  39. self.mask_before_denoising = True
  40. def get_pred_x0(self, x_in, x_out, sigma):
  41. ts = sigma.to(dtype=int)
  42. a_t = self.alphas[ts][:, None, None, None]
  43. sqrt_one_minus_at = (1 - a_t).sqrt()
  44. pred_x0 = (x_in - sqrt_one_minus_at * x_out) / a_t.sqrt()
  45. return pred_x0
  46. @property
  47. def inner_model(self):
  48. if self.model_wrap is None:
  49. denoiser = CompVisTimestepsVDenoiser if shared.sd_model.parameterization == "v" else CompVisTimestepsDenoiser
  50. self.model_wrap = denoiser(shared.sd_model)
  51. return self.model_wrap
  52. class CompVisSampler(sd_samplers_common.Sampler):
  53. def __init__(self, funcname, sd_model):
  54. super().__init__(funcname)
  55. self.eta_option_field = 'eta_ddim'
  56. self.eta_infotext_field = 'Eta DDIM'
  57. self.eta_default = 0.0
  58. self.model_wrap_cfg = CFGDenoiserTimesteps(self)
  59. self.model_wrap = self.model_wrap_cfg.inner_model
  60. def get_timesteps(self, p, steps):
  61. discard_next_to_last_sigma = self.config is not None and self.config.options.get('discard_next_to_last_sigma', False)
  62. if opts.always_discard_next_to_last_sigma and not discard_next_to_last_sigma:
  63. discard_next_to_last_sigma = True
  64. p.extra_generation_params["Discard penultimate sigma"] = True
  65. steps += 1 if discard_next_to_last_sigma else 0
  66. timesteps = torch.clip(torch.asarray(list(range(0, 1000, 1000 // steps)), device=devices.device) + 1, 0, 999)
  67. return timesteps
  68. def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None, image_conditioning=None):
  69. steps, t_enc = sd_samplers_common.setup_img2img_steps(p, steps)
  70. timesteps = self.get_timesteps(p, steps)
  71. timesteps_sched = timesteps[:t_enc]
  72. alphas_cumprod = shared.sd_model.alphas_cumprod
  73. sqrt_alpha_cumprod = torch.sqrt(alphas_cumprod[timesteps[t_enc]])
  74. sqrt_one_minus_alpha_cumprod = torch.sqrt(1 - alphas_cumprod[timesteps[t_enc]])
  75. xi = x * sqrt_alpha_cumprod + noise * sqrt_one_minus_alpha_cumprod
  76. if opts.img2img_extra_noise > 0:
  77. p.extra_generation_params["Extra noise"] = opts.img2img_extra_noise
  78. extra_noise_params = ExtraNoiseParams(noise, x, xi)
  79. extra_noise_callback(extra_noise_params)
  80. noise = extra_noise_params.noise
  81. xi += noise * opts.img2img_extra_noise * sqrt_alpha_cumprod
  82. extra_params_kwargs = self.initialize(p)
  83. parameters = inspect.signature(self.func).parameters
  84. if 'timesteps' in parameters:
  85. extra_params_kwargs['timesteps'] = timesteps_sched
  86. if 'is_img2img' in parameters:
  87. extra_params_kwargs['is_img2img'] = True
  88. self.model_wrap_cfg.init_latent = x
  89. self.last_latent = x
  90. self.sampler_extra_args = {
  91. 'cond': conditioning,
  92. 'image_cond': image_conditioning,
  93. 'uncond': unconditional_conditioning,
  94. 'cond_scale': p.cfg_scale,
  95. 's_min_uncond': self.s_min_uncond
  96. }
  97. samples = self.launch_sampling(t_enc + 1, lambda: self.func(self.model_wrap_cfg, xi, extra_args=self.sampler_extra_args, disable=False, callback=self.callback_state, **extra_params_kwargs))
  98. self.add_infotext(p)
  99. return samples
  100. def sample(self, p, x, conditioning, unconditional_conditioning, steps=None, image_conditioning=None):
  101. steps = steps or p.steps
  102. timesteps = self.get_timesteps(p, steps)
  103. extra_params_kwargs = self.initialize(p)
  104. parameters = inspect.signature(self.func).parameters
  105. if 'timesteps' in parameters:
  106. extra_params_kwargs['timesteps'] = timesteps
  107. self.last_latent = x
  108. self.sampler_extra_args = {
  109. 'cond': conditioning,
  110. 'image_cond': image_conditioning,
  111. 'uncond': unconditional_conditioning,
  112. 'cond_scale': p.cfg_scale,
  113. 's_min_uncond': self.s_min_uncond
  114. }
  115. samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, x, extra_args=self.sampler_extra_args, disable=False, callback=self.callback_state, **extra_params_kwargs))
  116. self.add_infotext(p)
  117. return samples
  118. sys.modules['modules.sd_samplers_compvis'] = sys.modules[__name__]
  119. VanillaStableDiffusionSampler = CompVisSampler # temp. compatibility with older extensions