4 Commits

Author	SHA1	Message	Date
Mr-Neutr0n	b68663426f	fix: preserve sign of denominator in clamp to avoid inverting gradient direction ... The previous .clamp(min=1e-6) on (sigma_ - sigma) flips the sign when the denominator is negative (which is the typical case since sigmas decrease monotonically). This would invert the target and cause training divergence. Use torch.sign(denom) * torch.clamp(denom.abs(), min=1e-6) instead, which prevents division by zero while preserving the correct sign.	2026-02-11 21:04:55 +05:30
Mr-Neutr0n	0e6976a0ae	fix: prevent division by zero in trajectory imitation loss at last step	2026-02-11 19:51:25 +05:30
Kared	8d0df403ca	fix wan i2v train bug	2026-01-27 03:55:36 +00:00
root	72af7122b3	DiffSynth-Studio 2.0 major update	2025-12-04 16:33:07 +08:00