Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-18 16:54:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix sign-destroying clamp in ER-SDE higher-order corrections

Browse Source 
The 2nd/3rd order derivative denominators (lambda_curr - prev_lambda)
are always negative since lambda decreases during denoising.
The .clamp(min=1e-10) forced these to a tiny positive value, causing
d_x0 to explode by ~1e18 and producing NaN output.

- Add _safe_div helper that returns zero for near-zero denominators
  while preserving the sign for normal values
- Replace .clamp(min=1e-10) with _safe_div in both 2nd and 3rd order
This commit is contained in:
CalamitousFelicitousness
2026-04-10 00:07:43 +01:00
parent 0b0382375a
commit 3f5059f310
1 changed files with 10 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/schedulers/scheduler_ersde.py
+10 -3
View File
@@ -306,6 +306,13 @@ class ERSDEScheduler(SchedulerMixin, ConfigMixin):
return torch.where(torch.abs(x) < eps, torch.zeros_like(x), x)
return 0.0 if abs(x) < eps else x
@staticmethod
def _safe_div(numerator, denominator, eps=1e-10):
"""Division with sign-preserving zero protection for lambda differences."""
if denominator.abs() < eps:
return torch.zeros_like(numerator)
return numerator / denominator
def _compute_fn(self, x):
"""Evaluate the noise scaling function on a scalar or tensor."""
if isinstance(x, (int, float)):
@@ -485,7 +492,7 @@ class ERSDEScheduler(SchedulerMixin, ConfigMixin):
elif effective_order == 2:
# 2nd order: add first derivative correction
d_x0 = (x0 - self.old_x0) / (lambda_curr - self.prev_lambda).clamp(min=1e-10)
d_x0 = self._safe_div(x0 - self.old_x0, lambda_curr - self.prev_lambda)
s_int = self._compute_integral(lambda_next, lambda_curr)
s_int = s_int.to(device)
delta_lambda = lambda_next - lambda_curr
@@ -494,8 +501,8 @@ class ERSDEScheduler(SchedulerMixin, ConfigMixin):
elif effective_order == 3:
# 3rd order: add second derivative correction
d_x0 = (x0 - self.old_x0) / (lambda_curr - self.prev_lambda).clamp(min=1e-10)
dd_x0 = 2.0 * (d_x0 - self.old_d_x0) / (lambda_curr - self.prev_prev_lambda).clamp(min=1e-10)
d_x0 = self._safe_div(x0 - self.old_x0, lambda_curr - self.prev_lambda)
dd_x0 = self._safe_div(2.0 * (d_x0 - self.old_d_x0), lambda_curr - self.prev_prev_lambda)
s_int = self._compute_integral(lambda_next, lambda_curr).to(device)
s_d_int = self._compute_derivative_integral(lambda_next, lambda_curr).to(device)
delta_lambda = lambda_next - lambda_curr