Update api/ltx_server_refactored.py

api/ltx_server_refactored.py

@@ -235,7 +235,7 @@ class VideoService:
         t0 = time.perf_counter()
         print("[INFO] Inicializando VideoService...")
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
-        self.config = self._load_config("ltxv-13b-0.9.8-distilled-fp8.yaml")
+        self.config = self._load_config("ltxv-13b-0.9.8-dev-fp8.yaml")
 
         self.pipeline, self.latent_upsampler = self._load_models_from_hub()
         self._move_models_to_device()
@@ -487,7 +487,7 @@ class VideoService:
         
         # --- Configuração de Seed e Diretórios ---
         used_seed = random.randint(0, 2**32 - 1) if seed is None else int(seed)
-        #seed_everything(used_seed)
+        self._seed_everething(used_seed)
         print(f"  - Usando Seed: {used_seed}")
     
         temp_dir = tempfile.mkdtemp(prefix="ltxv_low_")
@@ -495,12 +495,11 @@ class VideoService:
         results_dir = "/app/output"
         os.makedirs(results_dir, exist_ok=True)
     
-        # --- Cálculo de Dimensões e Frames ---
+        # --- Cálculo de Dimen.sões e Frames ---
         actual_num_frames = int(round(duration_secs * DEFAULT_FPS))
-        downscaled_height = height
-        downscaled_width = width
-        #self._calculate_downscaled_dims(height, width)
+        downscaled_height, downscaled_width = self._calculate_downscaled_dims(height, width)
 
+        num_frames = (actual_num_frames//8)+1
                     
         print(f"  - Frames: {actual_num_frames}, Duração: {duration_secs}s")
         print(f"  - Dimensões de Saída: {downscaled_height}x{downscaled_width}")
@@ -513,7 +512,7 @@ class VideoService:
                 "negative_prompt": negative_prompt,
                 "height": downscaled_height,
                 "width": downscaled_width,
-                "num_frames": (actual_num_frames//8)+1,
+                "num_frames": num_frames,
                 "frame_rate": int(DEFAULT_FPS),
                 "generator": torch.Generator(device=self.device).manual_seed(used_seed),
                 "output_type": "latent",
@@ -536,6 +535,9 @@ class VideoService:
             self._finalize()
 
             final_video_path, final_latents_path = self.refine_texture_only(
+                height=downscaled_height, 
+                width=downscaled_width,
+                num_frames=num_frames,
                 latents_path=tensor_path,
                 prompt=prompt,
                 negative_prompt=negative_prompt,
@@ -553,6 +555,7 @@ class VideoService:
     
     def refine_texture_only(
         self,
+        height, width, num_frames,
         latents_path: str,
         prompt: str,
         negative_prompt: str,
@@ -568,8 +571,8 @@ class VideoService:
         print("[INFO] Iniciando ETAPA 2: Refinamento de Textura...")
     
         # --- Configuração de Seed e Diretórios ---
-        used_seed = random.randint(0, 2**32 - 1) if seed is None else int(seed)
-        #seed_everything(used_seed)
+        used_seed = seed
+        self._seed_everething(used_seed)
         print(f"  - Usando Seed (consistente com Etapa 1): {used_seed}")
     
         temp_dir = tempfile.mkdtemp(prefix="ltxv_refine_single_")
@@ -592,10 +595,10 @@ class VideoService:
             second_pass_kwargs = {
                 "prompt": prompt,
                 "negative_prompt": negative_prompt,
-                "height": refine_height,
-                "width": refine_width,
+                "height": height,
+                "width": width,
                 "frame_rate": int(DEFAULT_FPS),
-                "num_frames": latents_to_refine.shape[2],
+                "num_frames": num_frames,
                 "latents": latents_to_refine,  # O tensor completo é passado aqui
                 "guidance_scale": float(guidance_scale),
                 "output_type": "latent",
@@ -607,7 +610,9 @@ class VideoService:
             print("  - Enviando tensor completo para a pipeline de refinamento...")
             final_latents = self.pipeline(**second_pass_kwargs).images
             print(f"  - Latentes refinados com shape: {final_latents.shape}")
-    
+
+            final_latents_filter =  adain_filter_latent(latents=final_latents, reference_latents=latents_to_refine)
+            
             # Decodifica os latentes refinados para pixels
             pixel_tensor = vae_manager_singleton.decode(final_latents, decode_timestep=float(self.config.get("decode_timestep", 0.05)))