diff options
Diffstat (limited to 'modules/sd_samplers.py')
| -rw-r--r-- | modules/sd_samplers.py | 261 |
1 files changed, 189 insertions, 72 deletions
diff --git a/modules/sd_samplers.py b/modules/sd_samplers.py index b58e810b..e904d860 100644 --- a/modules/sd_samplers.py +++ b/modules/sd_samplers.py @@ -1,32 +1,41 @@ -from collections import namedtuple
+from collections import namedtuple, deque
import numpy as np
+from math import floor
import torch
import tqdm
from PIL import Image
import inspect
import k_diffusion.sampling
+import torchsde._brownian.brownian_interval
import ldm.models.diffusion.ddim
import ldm.models.diffusion.plms
-from modules import prompt_parser, devices, processing
+from modules import prompt_parser, devices, processing, images, sd_vae_approx
from modules.shared import opts, cmd_opts, state
import modules.shared as shared
+from modules.script_callbacks import CFGDenoiserParams, cfg_denoiser_callback
SamplerData = namedtuple('SamplerData', ['name', 'constructor', 'aliases', 'options'])
samplers_k_diffusion = [
- ('Euler a', 'sample_euler_ancestral', ['k_euler_a'], {}),
+ ('Euler a', 'sample_euler_ancestral', ['k_euler_a', 'k_euler_ancestral'], {}),
('Euler', 'sample_euler', ['k_euler'], {}),
('LMS', 'sample_lms', ['k_lms'], {}),
('Heun', 'sample_heun', ['k_heun'], {}),
- ('DPM2', 'sample_dpm_2', ['k_dpm_2'], {}),
- ('DPM2 a', 'sample_dpm_2_ancestral', ['k_dpm_2_a'], {}),
+ ('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}),
+ ('DPM++ 2S a', 'sample_dpmpp_2s_ancestral', ['k_dpmpp_2s_a'], {}),
+ ('DPM++ 2M', 'sample_dpmpp_2m', ['k_dpmpp_2m'], {}),
+ ('DPM++ SDE', 'sample_dpmpp_sde', ['k_dpmpp_sde'], {}),
('DPM fast', 'sample_dpm_fast', ['k_dpm_fast'], {}),
('DPM adaptive', 'sample_dpm_adaptive', ['k_dpm_ad'], {}),
('LMS Karras', 'sample_lms', ['k_lms_ka'], {'scheduler': 'karras'}),
- ('DPM2 Karras', 'sample_dpm_2', ['k_dpm_2_ka'], {'scheduler': 'karras'}),
- ('DPM2 a Karras', 'sample_dpm_2_ancestral', ['k_dpm_2_a_ka'], {'scheduler': 'karras'}),
+ ('DPM2 Karras', 'sample_dpm_2', ['k_dpm_2_ka'], {'scheduler': 'karras', 'discard_next_to_last_sigma': True}),
+ ('DPM2 a Karras', 'sample_dpm_2_ancestral', ['k_dpm_2_a_ka'], {'scheduler': 'karras', 'discard_next_to_last_sigma': True}),
+ ('DPM++ 2S a Karras', 'sample_dpmpp_2s_ancestral', ['k_dpmpp_2s_a_ka'], {'scheduler': 'karras'}),
+ ('DPM++ 2M Karras', 'sample_dpmpp_2m', ['k_dpmpp_2m_ka'], {'scheduler': 'karras'}),
+ ('DPM++ SDE Karras', 'sample_dpmpp_sde', ['k_dpmpp_sde_ka'], {'scheduler': 'karras'}),
]
samplers_data_k_diffusion = [
@@ -40,16 +49,24 @@ all_samplers = [ SamplerData('DDIM', lambda model: VanillaStableDiffusionSampler(ldm.models.diffusion.ddim.DDIMSampler, model), [], {}),
SamplerData('PLMS', lambda model: VanillaStableDiffusionSampler(ldm.models.diffusion.plms.PLMSSampler, model), [], {}),
]
+all_samplers_map = {x.name: x for x in all_samplers}
samplers = []
samplers_for_img2img = []
+samplers_map = {}
-def create_sampler_with_index(list_of_configs, index, model):
- config = list_of_configs[index]
+def create_sampler(name, model):
+ if name is not None:
+ config = all_samplers_map.get(name, None)
+ else:
+ config = all_samplers[0]
+
+ assert config is not None, f'bad sampler name: {name}'
+
sampler = config.constructor(model)
sampler.config = config
-
+
return sampler
@@ -62,6 +79,12 @@ def set_samplers(): samplers = [x for x in all_samplers if x.name not in hidden]
samplers_for_img2img = [x for x in all_samplers if x.name not in hidden_img2img]
+ samplers_map.clear()
+ for sampler in all_samplers:
+ samplers_map[sampler.name.lower()] = sampler.name
+ for alias in sampler.aliases:
+ samplers_map[alias.lower()] = sampler.name
+
set_samplers()
@@ -71,6 +94,7 @@ sampler_extra_params = { 'sample_dpm_2': ['s_churn', 's_tmin', 's_tmax', 's_noise'],
}
+
def setup_img2img_steps(p, steps=None):
if opts.img2img_fix_steps or steps is not None:
steps = int((steps or p.steps) / min(p.denoising_strength, 0.999)) if p.denoising_strength > 0 else 0
@@ -82,14 +106,34 @@ def setup_img2img_steps(p, steps=None): return steps, t_enc
-def sample_to_image(samples):
- x_sample = processing.decode_first_stage(shared.sd_model, samples[0:1])[0]
+approximation_indexes = {"Full": 0, "Approx NN": 1, "Approx cheap": 2}
+
+
+def single_sample_to_image(sample, approximation=None):
+ if approximation is None:
+ approximation = approximation_indexes.get(opts.show_progress_type, 0)
+
+ if approximation == 2:
+ x_sample = sd_vae_approx.cheap_approximation(sample)
+ elif approximation == 1:
+ x_sample = sd_vae_approx.model()(sample.to(devices.device, devices.dtype).unsqueeze(0))[0].detach()
+ else:
+ x_sample = processing.decode_first_stage(shared.sd_model, sample.unsqueeze(0))[0]
+
x_sample = torch.clamp((x_sample + 1.0) / 2.0, min=0.0, max=1.0)
x_sample = 255. * np.moveaxis(x_sample.cpu().numpy(), 0, 2)
x_sample = x_sample.astype(np.uint8)
return Image.fromarray(x_sample)
+def sample_to_image(samples, index=0, approximation=None):
+ return single_sample_to_image(samples[index], approximation)
+
+
+def samples_to_image_grid(samples, approximation=None):
+ return images.image_grid([single_sample_to_image(sample, approximation) for sample in samples])
+
+
def store_latent(decoded):
state.current_latent = decoded
@@ -105,7 +149,8 @@ class InterruptedException(BaseException): class VanillaStableDiffusionSampler:
def __init__(self, constructor, sd_model):
self.sampler = constructor(sd_model)
- self.orig_p_sample_ddim = self.sampler.p_sample_ddim if hasattr(self.sampler, 'p_sample_ddim') else self.sampler.p_sample_plms
+ self.is_plms = hasattr(self.sampler, 'p_sample_plms')
+ self.orig_p_sample_ddim = self.sampler.p_sample_plms if self.is_plms else self.sampler.p_sample_ddim
self.mask = None
self.nmask = None
self.init_latent = None
@@ -117,6 +162,8 @@ class VanillaStableDiffusionSampler: self.config = None
self.last_latent = None
+ self.conditioning_key = sd_model.model.conditioning_key
+
def number_of_needed_noises(self, p):
return 0
@@ -136,6 +183,12 @@ class VanillaStableDiffusionSampler: if self.stop_at is not None and self.step > self.stop_at:
raise InterruptedException
+ # Have to unwrap the inpainting conditioning here to perform pre-processing
+ image_conditioning = None
+ if isinstance(cond, dict):
+ image_conditioning = cond["c_concat"][0]
+ cond = cond["c_crossattn"][0]
+ unconditional_conditioning = unconditional_conditioning["c_crossattn"][0]
conds_list, tensor = prompt_parser.reconstruct_multicond_batch(cond, self.step)
unconditional_conditioning = prompt_parser.reconstruct_cond_batch(unconditional_conditioning, self.step)
@@ -157,6 +210,12 @@ class VanillaStableDiffusionSampler: img_orig = self.sampler.model.q_sample(self.init_latent, ts)
x_dec = img_orig * self.mask + self.nmask * x_dec
+ # Wrap the image conditioning back up since the DDIM code can accept the dict directly.
+ # Note that they need to be lists because it just concatenates them later.
+ if image_conditioning is not None:
+ cond = {"c_concat": [image_conditioning], "c_crossattn": [cond]}
+ unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]}
+
res = self.orig_p_sample_ddim(x_dec, cond, ts, unconditional_conditioning=unconditional_conditioning, *args, **kwargs)
if self.mask is not None:
@@ -182,39 +241,52 @@ class VanillaStableDiffusionSampler: self.mask = p.mask if hasattr(p, 'mask') else None
self.nmask = p.nmask if hasattr(p, 'nmask') else None
- def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None):
- steps, t_enc = setup_img2img_steps(p, steps)
+ def adjust_steps_if_invalid(self, p, num_steps):
+ if (self.config.name == 'DDIM' and p.ddim_discretize == 'uniform') or (self.config.name == 'PLMS'):
+ valid_step = 999 / (1000 // num_steps)
+ if valid_step == floor(valid_step):
+ return int(valid_step) + 1
+
+ return num_steps
+ def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None, image_conditioning=None):
+ steps, t_enc = setup_img2img_steps(p, steps)
+ steps = self.adjust_steps_if_invalid(p, steps)
self.initialize(p)
- # existing code fails with certain step counts, like 9
- try:
- self.sampler.make_schedule(ddim_num_steps=steps, ddim_eta=self.eta, ddim_discretize=p.ddim_discretize, verbose=False)
- except Exception:
- self.sampler.make_schedule(ddim_num_steps=steps+1, ddim_eta=self.eta, ddim_discretize=p.ddim_discretize, verbose=False)
-
+ self.sampler.make_schedule(ddim_num_steps=steps, ddim_eta=self.eta, ddim_discretize=p.ddim_discretize, verbose=False)
x1 = self.sampler.stochastic_encode(x, torch.tensor([t_enc] * int(x.shape[0])).to(shared.device), noise=noise)
self.init_latent = x
+ self.last_latent = x
self.step = 0
- samples = self.launch_sampling(steps, lambda: self.sampler.decode(x1, conditioning, t_enc, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning))
+ # Wrap the conditioning models with additional image conditioning for inpainting model
+ if image_conditioning is not None:
+ conditioning = {"c_concat": [image_conditioning], "c_crossattn": [conditioning]}
+ unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]}
+
+
+ samples = self.launch_sampling(t_enc + 1, lambda: self.sampler.decode(x1, conditioning, t_enc, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning))
return samples
- def sample(self, p, x, conditioning, unconditional_conditioning, steps=None):
+ def sample(self, p, x, conditioning, unconditional_conditioning, steps=None, image_conditioning=None):
self.initialize(p)
self.init_latent = None
+ self.last_latent = x
self.step = 0
- steps = steps or p.steps
+ steps = self.adjust_steps_if_invalid(p, steps or p.steps)
- # existing code fails with certain step counts, like 9
- try:
- samples_ddim = self.launch_sampling(steps, lambda: self.sampler.sample(S=steps, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_T=x, eta=self.eta)[0])
- except Exception:
- samples_ddim = self.launch_sampling(steps, lambda: self.sampler.sample(S=steps+1, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_T=x, eta=self.eta)[0])
+ # Wrap the conditioning models with additional image conditioning for inpainting model
+ # dummy_for_plms is needed because PLMS code checks the first item in the dict to have the right shape
+ if image_conditioning is not None:
+ conditioning = {"dummy_for_plms": np.zeros((conditioning.shape[0],)), "c_crossattn": [conditioning], "c_concat": [image_conditioning]}
+ unconditional_conditioning = {"c_crossattn": [unconditional_conditioning], "c_concat": [image_conditioning]}
+
+ samples_ddim = self.launch_sampling(steps, lambda: self.sampler.sample(S=steps, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_T=x, eta=self.eta)[0])
return samples_ddim
@@ -228,7 +300,17 @@ class CFGDenoiser(torch.nn.Module): self.init_latent = None
self.step = 0
- def forward(self, x, sigma, uncond, cond, cond_scale):
+ 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 forward(self, x, sigma, uncond, cond, cond_scale, image_cond):
if state.interrupted or state.skipped:
raise InterruptedException
@@ -239,35 +321,37 @@ class CFGDenoiser(torch.nn.Module): repeats = [len(conds_list[i]) for i in range(batch_size)]
x_in = torch.cat([torch.stack([x[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [x])
+ image_cond_in = torch.cat([torch.stack([image_cond[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [image_cond])
sigma_in = torch.cat([torch.stack([sigma[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [sigma])
+ denoiser_params = CFGDenoiserParams(x_in, image_cond_in, sigma_in, state.sampling_step, state.sampling_steps)
+ cfg_denoiser_callback(denoiser_params)
+ x_in = denoiser_params.x
+ image_cond_in = denoiser_params.image_cond
+ sigma_in = denoiser_params.sigma
+
if tensor.shape[1] == uncond.shape[1]:
cond_in = torch.cat([tensor, uncond])
if shared.batch_cond_uncond:
- x_out = self.inner_model(x_in, sigma_in, cond=cond_in)
+ x_out = self.inner_model(x_in, sigma_in, cond={"c_crossattn": [cond_in], "c_concat": [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=cond_in[a:b])
+ x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond={"c_crossattn": [cond_in[a:b]], "c_concat": [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])
- x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=tensor[a:b])
+ x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond={"c_crossattn": [tensor[a:b]], "c_concat": [image_cond_in[a:b]]})
- x_out[-uncond.shape[0]:] = self.inner_model(x_in[-uncond.shape[0]:], sigma_in[-uncond.shape[0]:], cond=uncond)
+ x_out[-uncond.shape[0]:] = self.inner_model(x_in[-uncond.shape[0]:], sigma_in[-uncond.shape[0]:], cond={"c_crossattn": [uncond], "c_concat": [image_cond_in[-uncond.shape[0]:]]})
- 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)
+ 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
@@ -278,34 +362,63 @@ class CFGDenoiser(torch.nn.Module): class TorchHijack:
- def __init__(self, kdiff_sampler):
- self.kdiff_sampler = kdiff_sampler
+ 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.kdiff_sampler.randn_like
+ return self.randn_like
if hasattr(torch, item):
return getattr(torch, item)
raise AttributeError("'{}' object has no attribute '{}'".format(type(self).__name__, item))
+ def randn_like(self, x):
+ if self.sampler_noises:
+ noise = self.sampler_noises.popleft()
+ if noise.shape == x.shape:
+ return noise
+
+ if x.device.type == 'mps':
+ return torch.randn_like(x, device=devices.cpu).to(x.device)
+ else:
+ return torch.randn_like(x)
+
+
+# MPS fix for randn in torchsde
+def torchsde_randn(size, dtype, device, seed):
+ if device.type == 'mps':
+ generator = torch.Generator(devices.cpu).manual_seed(int(seed))
+ return torch.randn(size, dtype=dtype, device=devices.cpu, generator=generator).to(device)
+ else:
+ generator = torch.Generator(device).manual_seed(int(seed))
+ return torch.randn(size, dtype=dtype, device=device, generator=generator)
+
+
+torchsde._brownian.brownian_interval._randn = torchsde_randn
+
class KDiffusionSampler:
def __init__(self, funcname, sd_model):
- self.model_wrap = k_diffusion.external.CompVisDenoiser(sd_model, quantize=shared.opts.enable_quantization)
+ 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.sampler_noise_index = 0
self.stop_at = None
self.eta = None
self.default_eta = 1.0
self.config = None
self.last_latent = None
+ self.conditioning_key = sd_model.model.conditioning_key
+
def callback_state(self, d):
step = d['i']
latent = d["denoised"]
@@ -330,26 +443,13 @@ class KDiffusionSampler: def number_of_needed_noises(self, p):
return p.steps
- def randn_like(self, x):
- noise = self.sampler_noises[self.sampler_noise_index] if self.sampler_noises is not None and self.sampler_noise_index < len(self.sampler_noises) else None
-
- if noise is not None and x.shape == noise.shape:
- res = noise
- else:
- res = torch.randn_like(x)
-
- self.sampler_noise_index += 1
- return res
-
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.step = 0
- self.sampler_noise_index = 0
self.eta = p.eta or opts.eta_ancestral
- if self.sampler_noises is not None:
- k_diffusion.sampling.torch = TorchHijack(self)
+ 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:
@@ -361,16 +461,26 @@ class KDiffusionSampler: return extra_params_kwargs
- def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None):
- steps, t_enc = setup_img2img_steps(p, steps)
-
+ def get_sigmas(self, p, steps):
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':
- sigmas = k_diffusion.sampling.get_sigmas_karras(n=steps, sigma_min=0.1, sigma_max=10, device=shared.device)
+ 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 self.config is not None and self.config.options.get('discard_next_to_last_sigma', False):
+ sigmas = torch.cat([sigmas[:-2], sigmas[-1:]])
+
+ return sigmas
+
+ def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None, image_conditioning=None):
+ steps, t_enc = setup_img2img_steps(p, steps)
+
+ sigmas = self.get_sigmas(p, steps)
+
sigma_sched = sigmas[steps - t_enc - 1:]
xi = x + noise * sigma_sched[0]
@@ -388,20 +498,21 @@ class KDiffusionSampler: extra_params_kwargs['sigmas'] = sigma_sched
self.model_wrap_cfg.init_latent = x
+ self.last_latent = x
- samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, xi, extra_args={'cond': conditioning, 'uncond': unconditional_conditioning, 'cond_scale': p.cfg_scale}, disable=False, callback=self.callback_state, **extra_params_kwargs))
+ samples = self.launch_sampling(t_enc + 1, lambda: self.func(self.model_wrap_cfg, xi, extra_args={
+ 'cond': conditioning,
+ 'image_cond': image_conditioning,
+ 'uncond': unconditional_conditioning,
+ 'cond_scale': p.cfg_scale
+ }, disable=False, callback=self.callback_state, **extra_params_kwargs))
return samples
- def sample(self, p, x, conditioning, unconditional_conditioning, steps=None):
+ def sample(self, p, x, conditioning, unconditional_conditioning, steps=None, image_conditioning = None):
steps = steps or p.steps
- 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':
- sigmas = k_diffusion.sampling.get_sigmas_karras(n=steps, sigma_min=0.1, sigma_max=10, device=shared.device)
- else:
- sigmas = self.model_wrap.get_sigmas(steps)
+ sigmas = self.get_sigmas(p, steps)
x = x * sigmas[0]
@@ -414,7 +525,13 @@ class KDiffusionSampler: else:
extra_params_kwargs['sigmas'] = sigmas
- samples = self.launch_sampling(steps, lambda: self.func(self.model_wrap_cfg, x, extra_args={'cond': conditioning, 'uncond': unconditional_conditioning, 'cond_scale': p.cfg_scale}, disable=False, callback=self.callback_state, **extra_params_kwargs))
+ 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
+ }, disable=False, callback=self.callback_state, **extra_params_kwargs))
return samples
|