Fix self-correcting scheduler infinite loop and strict zero check

- Add gap-size guard to prevent infinite insertions when invariants
  trigger repeatedly on the same step
- Replace strict sigmas[-1] == 0.0 check with tolerance-based clamp
  (fixes KSamplerAdvanced / sliced-schedule compatibility)

Signed-off-by: Galih Tama <galpt@v.recipes>

Author: galpt

infinity_diffusion.py

@@ -219,8 +219,11 @@ def sample(
         """
         if sigmas.ndim != 1:
             raise ValueError(f"sigmas must be 1-D, got shape {sigmas.shape}")
-        if sigmas[-1] != 0.0:
-            raise ValueError("last element of sigmas must be 0")
+        # Clamp near-zero last sigma to exactly 0 for compatibility with
+        # KSamplerAdvanced and similar nodes that may pass a sliced schedule.
+        if sigmas[-1].abs() > 1e-6:
+            sigmas = sigmas.clone()
+            sigmas[-1] = 0.0
         if sigmas.numel() < 2:
             raise ValueError("sigmas must have at least 2 elements")
 
@@ -277,16 +280,19 @@ def sample(
             reversed_dir = cos_sim < 0.0
 
             # Self-correcting scheduler: if an invariant triggered, insert a step
-            # and redo this step for finer resolution.
-            if clamped or reversed_dir:
-                mid = (sigmas_list[i] + sigmas_list[i + 1]) * 0.5
-                sigmas_list.insert(i + 1, mid)
-                continue  # redo `i` with finer resolution
+            # and redo this step for finer resolution.  The gap check prevents
+            # infinite loops when the step is already too small to split further.
+            if (clamped or reversed_dir) and i < len(sigmas_list) - 1:
+                current_gap = (sigmas_list[i] - sigmas_list[i + 1]).abs().item()
+                if current_gap > 1e-6:
+                    mid = (sigmas_list[i] + sigmas_list[i + 1]) * 0.5
+                    sigmas_list.insert(i + 1, mid)
+                    continue
 
             # Fallback: Euler step if both invariants fail
             if clamped and reversed_dir:
                 correction = torch.zeros_like(raw_correction)
-            elif reversed_dir:
+            elif inv_reversed:
                 correction = raw_correction * 0.5
             else:
                 correction = raw_correction