Refactor video and audio latent generation in generate_video and generate_video_with_audio

- Removed direct initialization of latents with random noise, replacing it with a conditional approach based on I2V (Image-to-Video) conditioning.
- Introduced a structured flow for applying noise during the latent state creation, enhancing the conditioning process for both video and audio.
- Updated the noise application logic to ensure proper handling of conditioned and unconditioned frames in both stages of video generation.
- Improved code clarity and maintainability by consolidating latent shape definitions and restructuring noise application logic.

mlx_video/generate.py

@@ -332,8 +332,6 @@ def generate_video(
     # Stage 1: Generate at half resolution
     print(f"{Colors.YELLOW}⚡ Stage 1: Generating at {width//2}x{height//2} (8 steps)...{Colors.RESET}")
     mx.random.seed(seed)
-    latents = mx.random.normal((1, 128, latent_frames, stage1_h, stage1_w))
-    mx.eval(latents)
 
     positions = create_position_grid(1, latent_frames, stage1_h, stage1_w)
     mx.eval(positions)
@@ -341,10 +339,12 @@ def generate_video(
     # Apply I2V conditioning if provided
     state1 = None
     if is_i2v and stage1_image_latent is not None:
-        # Create state with conditioning
+        # PyTorch flow: create zeros -> apply conditioning -> apply noiser
+        # Create initial state with zeros
+        latent_shape = (1, 128, latent_frames, stage1_h, stage1_w)
         state1 = LatentState(
-            latent=latents,
-            clean_latent=mx.zeros_like(latents),
+            latent=mx.zeros(latent_shape),
+            clean_latent=mx.zeros(latent_shape),
             denoise_mask=mx.ones((1, 1, latent_frames, 1, 1)),
         )
         conditioning = VideoConditionByLatentIndex(
@@ -353,8 +353,23 @@ def generate_video(
             strength=image_strength,
         )
         state1 = apply_conditioning(state1, [conditioning])
+
+        # Apply noiser: latent = noise * (mask * noise_scale) + latent * (1 - mask * noise_scale)
+        # For Stage 1, noise_scale = 1.0 (first sigma)
+        noise = mx.random.normal(latent_shape)
+        noise_scale = STAGE_1_SIGMAS[0]  # 1.0
+        scaled_mask = state1.denoise_mask * noise_scale
+        state1 = LatentState(
+            latent=noise * scaled_mask + state1.latent * (1.0 - scaled_mask),
+            clean_latent=state1.clean_latent,
+            denoise_mask=state1.denoise_mask,
+        )
         latents = state1.latent
         mx.eval(latents)
+    else:
+        # T2V: just use random noise
+        latents = mx.random.normal((1, 128, latent_frames, stage1_h, stage1_w))
+        mx.eval(latents)
 
     latents = denoise(latents, positions, text_embeddings, transformer, STAGE_1_SIGMAS, verbose=verbose, state=state1)
 
@@ -379,17 +394,12 @@ def generate_video(
     positions = create_position_grid(1, latent_frames, stage2_h, stage2_w)
     mx.eval(positions)
 
-    # Add noise for refinement
-    noise_scale = STAGE_2_SIGMAS[0]
-    noise = mx.random.normal(latents.shape)
-    latents = noise * noise_scale + latents * (1 - noise_scale)
-    mx.eval(latents)
-
     # Apply I2V conditioning for stage 2 if provided
     state2 = None
     if is_i2v and stage2_image_latent is not None:
+        # PyTorch flow: start with upscaled latent -> apply conditioning -> apply noiser
         state2 = LatentState(
-            latent=latents,
+            latent=latents,  # Start with upscaled latent
             clean_latent=mx.zeros_like(latents),
             denoise_mask=mx.ones((1, 1, latent_frames, 1, 1)),
         )
@@ -399,8 +409,26 @@ def generate_video(
             strength=image_strength,
         )
         state2 = apply_conditioning(state2, [conditioning])
+
+        # Apply noiser: latent = noise * (mask * noise_scale) + latent * (1 - mask * noise_scale)
+        # For Stage 2, noise_scale = stage_2_sigmas[0]
+        # Conditioned frames (mask=0) keep image latent, unconditioned get partial noise
+        noise = mx.random.normal(latents.shape)
+        noise_scale = STAGE_2_SIGMAS[0]
+        scaled_mask = state2.denoise_mask * noise_scale
+        state2 = LatentState(
+            latent=noise * scaled_mask + state2.latent * (1.0 - scaled_mask),
+            clean_latent=state2.clean_latent,
+            denoise_mask=state2.denoise_mask,
+        )
         latents = state2.latent
         mx.eval(latents)
+    else:
+        # T2V: add noise to all frames for refinement
+        noise_scale = STAGE_2_SIGMAS[0]
+        noise = mx.random.normal(latents.shape)
+        latents = noise * noise_scale + latents * (1 - noise_scale)
+        mx.eval(latents)
 
     latents = denoise(latents, positions, text_embeddings, transformer, STAGE_2_SIGMAS, verbose=verbose, state=state2)