fix: video RoPE clock uses the render fps (hardcoded 24 at 25fps renders = 4%/s AV rope skew = the ~10s lip-sync drift)

libs/ltx_distillation/models/ltx_wrapper.py

@@ -231,8 +231,13 @@ class LTX2DiffusionWrapper(nn.Module):
             causal_fix=True,
         ).float()
 
-        # Convert temporal dimension from frames to seconds (divide by fps=24)
-        # This matches VideoLatentTools.create_initial_state
+        # Convert temporal dimension from frames to seconds. self.VIDEO_FPS
+        # MUST be the actual render fps - the JoyEcho nodes stamp it as an
+        # instance attribute before generation. The class default of 24.0 with
+        # 25fps renders skewed video rope-time 4% fast against audio's
+        # true-seconds rope: a growing mouth-ahead-of-audio drift that crossed
+        # visibility at ~9.6s into every shot (the "10s lip-sync cliff",
+        # diagnosed 2026-07-23).
         pixel_coords[:, 0, ...] = pixel_coords[:, 0, ...] / self.VIDEO_FPS
 
         if downscale_factor != 1:

nodes.py

@@ -1287,6 +1287,12 @@ class JoyEcho_Generate:
         generator.latent_height = video_height // 32
         generator.latent_width = video_width // 32
         generator.video_frame_seqlen = generator.latent_height * generator.latent_width
+        # RENDER fps must drive the video RoPE clock (wrapper class default is
+        # 24.0). At 25fps the hardcoded 24 skewed video rope-time 4% fast vs
+        # audio's true-seconds rope -> ~40ms/s growing mouth-ahead-of-audio
+        # drift, crossing visibility at ~9.6s into every shot (the "10s lip
+        # sync cliff", 2026-07-23).
+        generator.VIDEO_FPS = float(video_fps)
 
         # Compute latent shapes
         video_shape, audio_shape = compute_latent_shapes(
@@ -1890,6 +1896,7 @@ class JoyEcho_Generate:
         generator.latent_height = th // 32
         generator.latent_width = tw // 32
         generator.video_frame_seqlen = generator.latent_height * generator.latent_width
+        generator.VIDEO_FPS = float(fps)  # refine RoPE clock = render fps too
 
         offl = None
         if sequential_offload:
@@ -2228,6 +2235,8 @@ class JoyEcho_SingleShotGenerate:
         generator.latent_height = video_height // 32
         generator.latent_width = video_width // 32
         generator.video_frame_seqlen = generator.latent_height * generator.latent_width
+        # render fps drives the video RoPE clock (see multishot Generate note)
+        generator.VIDEO_FPS = float(video_fps)
 
         # Compute latent shapes
         video_shape, audio_shape = compute_latent_shapes(