[bugfix] LTX2: honor video_position_offset_sec in the DiT (#1422)

2026-06-01 16:59:15 -07:00
parent 89fcf08378
commit 3f24578139
1 changed files with 5 additions and 1 deletions
@@ -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: