Spaces:
Running on Zero
Running on Zero
File size: 11,833 Bytes
b701455 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 | """K-sampler utilities for diffusion models."""
import collections
import logging
import numpy as np
import scipy
import torch
from src.sample import sampling_util
def calculate_start_end_timesteps(model: torch.nn.Module, conds: list) -> None:
"""Calculate start/end timesteps for conditions."""
s = model.model_sampling
for t in range(len(conds)):
x = conds[t]
ts, te = x.get("start_percent"), x.get("end_percent")
if ts is not None or te is not None:
n = x.copy()
if ts is not None: n["timestep_start"] = s.percent_to_sigma(ts)
if te is not None: n["timestep_end"] = s.percent_to_sigma(te)
conds[t] = n
def pre_run_control(model: torch.nn.Module, conds: list) -> None:
"""Pre-run control for conditions."""
s = model.model_sampling
for x in conds:
if "control" in x:
x["control"].pre_run(model, lambda a: s.percent_to_sigma(a))
def apply_empty_x_to_equal_area(conds: list, uncond: list, name: str, uncond_fill_func: callable) -> None:
"""Apply empty x to equal area."""
cond_cnets, cond_other = [], []
uncond_cnets, uncond_other = [], []
for t, x in enumerate(conds):
if "area" not in x:
(cond_cnets if name in x and x[name] else cond_other).append((x[name], None) if name in x and x[name] else (x, t))
for t, x in enumerate(uncond):
if "area" not in x:
(uncond_cnets if name in x and x[name] else uncond_other).append((x[name], None) if name in x and x[name] else (x, t))
if uncond_cnets: return
for i, _ in enumerate(cond_cnets):
temp = uncond_other[i % len(uncond_other)]
n = temp[0].copy()
n[name] = uncond_fill_func([c[0] for c in cond_cnets if c[1] is None], i)
if temp[1] is not None: uncond[temp[1]] = n
else: uncond.append(n)
CondObj = collections.namedtuple("cond_obj", ["input_x", "mult", "conditioning", "area", "control", "patches", "batch_indices"])
def get_area_and_mult(conds: dict, x_in: torch.Tensor, timestep_in: int) -> CondObj:
"""Get area and multiplier from conditions."""
x_shape, device = x_in.shape, x_in.device
area = (x_shape[2], x_shape[3], 0, 0)
batch_indices = conds.get("batch_index")
if isinstance(batch_indices, int): batch_indices = [batch_indices]
area_h, area_w = max(0, min(int(area[0]), x_shape[2])), max(0, min(int(area[1]), x_shape[3]))
area = (area_h, area_w, 0, 0)
if batch_indices is None:
input_x = x_in[:, :, :area_h, :area_w]
else:
try:
mapped = [(int(b) if b >= 0 else x_shape[0] + int(b)) for b in batch_indices]
valid = [b for b in mapped if 0 <= b < x_shape[0]]
if not valid:
batch_indices = None
input_x = x_in[:, :, :area_h, :area_w]
else:
input_x = x_in[torch.tensor(valid, dtype=torch.long, device=device), :, :area_h, :area_w]
except Exception:
batch_indices = None
input_x = x_in[:, :, :area_h, :area_w]
mult = torch.ones_like(input_x)
batch_size = x_shape[0] if batch_indices is None else len(batch_indices)
# Handle mock objects in tests
if not isinstance(batch_size, int):
try:
temp = int(batch_size)
if isinstance(temp, int):
batch_size = temp
else:
batch_size = 1
except Exception:
batch_size = 1
if not isinstance(device, (torch.device, str)):
from src.Device import Device
device = Device.get_torch_device()
conditioning = {c: conds["model_conds"][c].process_cond(batch_size=batch_size, device=device, area=area)
for c in conds["model_conds"]}
return CondObj(input_x, mult, conditioning, area, conds.get("control"), None, batch_indices)
def normal_scheduler(model_sampling, steps: int, sgm: bool = False, floor: bool = False) -> torch.FloatTensor:
"""Create normal noise scheduler."""
s = model_sampling
timesteps = torch.linspace(s.timestep(s.sigma_max), s.timestep(s.sigma_min), steps, device=s.sigmas.device)
return torch.cat([s.sigma(timesteps), s.sigmas.new_zeros([1])]).cpu().float()
def simple_scheduler(model_sampling, steps: int) -> torch.FloatTensor:
"""Create simple noise scheduler."""
s = model_sampling
if steps <= 0: return torch.FloatTensor([0.0])
indices = (torch.arange(steps, device=s.sigmas.device) * len(s.sigmas) / steps).long()
sigs = s.sigmas.flip(0)[indices]
return torch.cat([sigs, sigs.new_zeros([1])]).cpu().float()
def beta_scheduler(model_sampling, steps, alpha=0.6, beta=0.6) -> torch.FloatTensor:
"""Create beta distribution noise scheduler."""
total = len(model_sampling.sigmas) - 1
ts = scipy.stats.beta.ppf(1 - np.linspace(0, 1, steps, endpoint=False), alpha, beta)
ts_indices = np.rint(ts * total).astype(np.int32)
unique_ts, indices = np.unique(ts_indices, return_index=True)
ordered = unique_ts[np.argsort(indices)]
sigs = model_sampling.sigmas[torch.from_numpy(ordered).to(model_sampling.sigmas.device, torch.long)]
return torch.cat([sigs, sigs.new_zeros([1])]).cpu().float()
def _compute_flux2_mu(image_seq_len: int, num_steps: int) -> float:
"""Compute empirical mu for Flux2 scheduler (matches ComfyUI exactly).
This resolution-dependent mu calculation is critical for Flux2 quality.
"""
a1, b1 = 8.73809524e-05, 1.89833333
a2, b2 = 0.00016927, 0.45666666
if image_seq_len > 4300:
return a2 * image_seq_len + b2
m_200 = a2 * image_seq_len + b2
m_10 = a1 * image_seq_len + b1
a = (m_200 - m_10) / 190.0
b = m_200 - 200.0 * a
return a * num_steps + b
def _flux2_time_shift(t: torch.Tensor, mu: float, sigma: float = 1.0) -> torch.Tensor:
"""Generalized time SNR shift for Flux2 (matches ComfyUI exactly)."""
import math
return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
def flux2_scheduler(steps: int, width: int, height: int) -> torch.FloatTensor:
"""Create Flux2 noise scheduler (matches ComfyUI Flux2Scheduler exactly).
This scheduler dynamically computes mu based on image resolution and steps,
which is critical for Flux2 image quality.
Args:
steps: Number of sampling steps
width: Image width in pixels
height: Image height in pixels
Returns:
Sigmas tensor of shape (steps + 1,) ending with 0
"""
# Calculate sequence length (number of 16x16 patches)
seq_len = round((width * height) / (16 * 16))
# Compute resolution/steps-dependent mu
mu = _compute_flux2_mu(seq_len, steps)
# Create timesteps from 1 to 0 (inclusive)
timesteps = torch.linspace(1, 0, steps + 1)
# Apply time shift - avoid division by zero at t=0
sigmas = torch.zeros_like(timesteps)
mask = timesteps > 0
sigmas[mask] = _flux2_time_shift(timesteps[mask], mu)
sigmas[~mask] = 0.0 # t=0 maps to sigma=0
return sigmas.cpu().float()
def calculate_sigmas(model_sampling, scheduler_name: str, steps: int,
width: int = None, height: int = None, is_flux2: bool = False) -> torch.Tensor:
"""Calculate sigmas for scheduler.
For Flux2 models, use the resolution-aware Flux2Scheduler when width/height are provided.
This matches ComfyUI's behavior and is critical for image quality.
"""
# Robust Flux2 detection if flag not set
if not is_flux2 and model_sampling:
cls_name = model_sampling.__class__.__name__
if "ModelSamplingFlux2" in cls_name:
is_flux2 = True
# Handle mock objects in tests
if not isinstance(steps, int):
try:
steps = int(steps)
except Exception:
steps = 20
# For Flux2 with resolution info, use the dedicated Flux2 scheduler (matches ComfyUI)
if is_flux2 and width is not None and height is not None:
return flux2_scheduler(steps, width, height)
if scheduler_name == "karras":
return sampling_util.get_sigmas_karras(steps, float(model_sampling.sigma_min), float(model_sampling.sigma_max))
elif scheduler_name == "normal":
return normal_scheduler(model_sampling, steps)
elif scheduler_name == "simple":
return simple_scheduler(model_sampling, steps)
elif scheduler_name == "beta":
return beta_scheduler(model_sampling, steps)
elif scheduler_name in ["ays", "ays_sd15", "ays_sdxl"]:
from src.sample import ays_scheduler as ays
model_type = {"ays_sdxl": "SDXL", "ays_sd15": "SD15"}.get(scheduler_name)
if not model_type:
try:
# Robust detection based on class name or config flags
cls_name = model_sampling.__class__.__name__.lower()
if "flux" in cls_name:
model_type = "FLUX"
else:
config = getattr(model_sampling, 'model_config', None)
if config and getattr(config, 'is_flux', False):
model_type = "FLUX"
elif config and getattr(config, 'uses_dual_clip', False):
model_type = "SDXL"
else:
# Fallback to context_dim check
unet_config = getattr(config, 'unet_config', {})
model_type = "SDXL" if unet_config.get('context_dim', 0) == 2048 else "SD15"
except:
model_type = "SD15"
return ays.ays_scheduler(model_sampling, steps, model_type)
logging.error(f"Invalid scheduler: {scheduler_name}")
return None
def prepare_noise(latent_image: torch.Tensor, seed: int, noise_inds: list = None,
seeds_per_sample: list | None = None) -> torch.Tensor:
"""Prepare noise for latent image.
NOTE: Noise is generated on CPU for reproducibility across devices (matching ComfyUI behavior).
Using a GPU generator produces different random values than CPU even with the same seed.
"""
target_device = latent_image.device
if seeds_per_sample is not None:
sps = np.array(seeds_per_sample)
if sps.shape[0] != latent_image.size(0):
raise ValueError("seeds_per_sample length must match latent batch size")
unique_seeds, inverse = np.unique(sps, return_inverse=True)
noises = []
for us in unique_seeds:
g = torch.Generator(device="cpu")
g.manual_seed(int(us))
# Generate on CPU for reproducibility, then move to target device
noises.append(torch.randn([1] + list(latent_image.size())[1:], dtype=latent_image.dtype,
layout=latent_image.layout, generator=g, device="cpu").to(target_device))
return torch.cat([noises[i] for i in inverse], axis=0)
generator = torch.Generator(device="cpu")
generator.manual_seed(seed)
if noise_inds is None:
# Generate on CPU for reproducibility (matches ComfyUI), then move to target device
return torch.randn(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout,
generator=generator, device="cpu").to(target_device)
unique_inds, inverse = np.unique(noise_inds, return_inverse=True)
noises = []
for i in range(unique_inds[-1] + 1):
noise = torch.randn([1] + list(latent_image.size())[1:], dtype=latent_image.dtype,
layout=latent_image.layout, generator=generator, device="cpu").to(target_device)
if i in unique_inds: noises.append(noise)
return torch.cat([noises[i] for i in inverse], axis=0)
|