From 3f245781391922d8270ec3efcbcf130d8bd115bf Mon Sep 17 00:00:00 2001 From: Kaiqin Kong Date: Mon, 1 Jun 2026 16:59:15 -0700 Subject: [PATCH] [bugfix] LTX2: honor `video_position_offset_sec` in the DiT (#1422) --- fastvideo/models/dits/ltx2.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/fastvideo/models/dits/ltx2.py b/fastvideo/models/dits/ltx2.py index 1cda58a93..aba91df56 100644 --- a/fastvideo/models/dits/ltx2.py +++ b/fastvideo/models/dits/ltx2.py @@ -2907,6 +2907,7 @@ class LTX2Transformer3DModel(BaseDiT): skip_cross_modal_attn: bool = False, skip_video_self_attn_blocks: list[int] | None = None, skip_audio_self_attn_blocks: list[int] | None = None, + video_position_offset_sec: float = 0.0, **kwargs, ) -> torch.Tensor: if isinstance(encoder_hidden_states, list): @@ -2959,7 +2960,10 @@ class LTX2Transformer3DModel(BaseDiT): DEFAULT_LTX2_SCALE_FACTORS, fps=fps, causal_fix=True, - ).to(hidden_states.dtype) + ) + if video_position_offset_sec: + positions[:, 0, ...] = positions[:, 0, ...] + float(video_position_offset_sec) + positions = positions.to(hidden_states.dtype) # Pad positions to match padded sequence length for SP cross-attention if sp_world_size > 1 and video_padded_seq_len > video_original_seq_len: