Update api/ltx_server.py

api/ltx_server.py

@@ -754,27 +754,22 @@ class VideoService:
                     # --- ETAPA 1: GERAÇÃO BASE (FIRST PASS) ---
                     print("\n--- INICIANDO ETAPA 1: GERAÇÃO BASE (FIRST PASS) ---")
                     t_pass1 = time.perf_counter()
-                    
-                    first_pass_config = self.config.get("first_pass", {}).copy()
-                    downscale_factor = self.config.get("downscale_factor", 0.666)
 
-                    unrounded_width = width_padded * downscale_factor
-                    unrounded_height = height_padded * downscale_factor
-                    
-                    # Sanitiza as dimensões para serem divisíveis pelo fator do VAE (geralmente 8)
-                    # Usamos o mesmo divisor da UI para consistência.
-                    divisor = 8
-                    downscaled_width = int(round(unrounded_width / divisor)) * divisor
-                    downscaled_height = int(round(unrounded_height / divisor)) * divisor
-
-                    downscaled_width = max(divisor, downscaled_width)
-                    downscaled_height = max(divisor, downscaled_height)
-                    first_pass_kwargs.update(first_pass_config)
-                    
-                    
-                    print(f"[DEBUG] Dimensões do First Pass: Calculado ({unrounded_width:.0f}x{unrounded_height:.0f}) -> Sanitizado ({downscaled_width}x{downscaled_height})")
+                    first_pass_config = self.config.get("first_pass", {}).copy()
+                    downscale_factor = self.config.get("downscale_factor", 0.6666666)
+                    vae_scale_factor = self.pipeline.vae_scale_factor # Geralmente 8
                     
-                    first_pass_kwargs = call_kwargs.copy()
+                    # --- <INÍCIO DA LÓGICA DE CÁLCULO EXATA> ---
+                    # Replica a fórmula da LTXMultiScalePipeline
+                    x_width = int(width_padded * downscale_factor)
+                    downscaled_width = x_width - (x_width % vae_scale_factor)
+                    x_height = int(height_padded * downscale_factor)
+                    downscaled_height = x_height - (x_height % vae_scale_factor)
+                    print(f"[DEBUG] First Pass Dims: Original Pad ({width_padded}x{height_padded}) -> Downscaled ({downscaled_width}x{downscaled_height})")
+                    # --- <FIM DA LÓGICA DE CÁLCULO EXATA> ---
+
+                    first_pass_kwargs = base_call_kwargs.copy()
+    
                     first_pass_kwargs.update({
                         "output_type": "latent",
                         "width": downscaled_width,
@@ -799,22 +794,23 @@ class VideoService:
                     del base_latents; gc.collect(); torch.cuda.empty_cache()
 
 
-
-                    
-                    # --- ETAPA 3: REFINAMENTO DE TEXTURA (SECOND PASS) ---
+                    # # --- ETAPA 3: REFINAMENTO DE TEXTURA (SECOND PASS) ---
                     print("\n--- INICIANDO ETAPA 3: REFINAMENTO DE TEXTURA (SECOND PASS) ---")
-                    t_pass2 = time.perf_counter()
                     
                     second_pass_config = self.config.get("second_pass", {}).copy()
-                    second_pass_kwargs.update(second_pass_config)
-                    
+                    # --- <INÍCIO DA LÓGICA DE CÁLCULO EXATA PARA SECOND PASS> ---
+                    # Usa as dimensões da primeira passagem dobradas, como na pipeline original
+                    second_pass_width = downscaled_width * 2
+                    second_pass_height = downscaled_height * 2
+                    print(f"[DEBUG] Second Pass Dims: Target ({second_pass_width}x{second_pass_height})")
+                    # --- <FIM DA LÓGICA DE CÁLCULO EXATA> --- 
                     
-                    second_pass_kwargs = call_kwargs.copy()
+                    second_pass_kwargs = base_call_kwargs.copy()
                     second_pass_kwargs.update({
                         "output_type": "latent",
-                        "width": width_padded,
-                        "height": height_padded,
-                        "latents": upsampled_latents,
+                        "width": second_pass_width,
+                        "height": second_pass_height,
+                        "latents": upsampled_latents, # O tensor upscaled
                         "guidance_scale": float(guidance_scale),
                         **second_pass_config
                     })