Update api/ltx_server.py

api/ltx_server.py

@@ -577,39 +577,103 @@ class VideoService:
         latents = None
         multi_scale_pipeline = None
 
+        # ltx_server.py (dentro da função generate)
+
         try:
             if improve_texture:
                 if not self.latent_upsampler:
                     raise ValueError("Upscaler espacial não carregado.")
-                print("[DEBUG] Multi-escala: construindo pipeline...")
-                multi_scale_pipeline = LTXMultiScalePipeline(self.pipeline, self.latent_upsampler)
+                
+                print("[DEBUG] Multi-escala: Iniciando Passo 1 (geração de latentes base).")
+                
+                # 1. Configurar e executar o primeiro passo
                 first_pass_args = self.config.get("first_pass", {}).copy()
-                first_pass_args["guidance_scale"] = float(guidance_scale)
-                second_pass_args = self.config.get("second_pass", {}).copy()
-                second_pass_args["guidance_scale"] = float(guidance_scale)
-
-                multi_scale_call_kwargs = call_kwargs.copy()
-                multi_scale_call_kwargs.update(
-                    {
-                        "downscale_factor": self.config["downscale_factor"],
-                        "first_pass": first_pass_args,
-                        "second_pass": second_pass_args,
-                    }
-                )
-                print("[DEBUG] Chamando multi_scale_pipeline...")
-                t_ms = time.perf_counter()
+                first_pass_kwargs = call_kwargs.copy()
+                first_pass_kwargs.update({
+                    "guidance_scale": float(guidance_scale),
+                    "stg_scale": first_pass_args.get("stg_scale"),
+                    "rescaling_scale": first_pass_args.get("rescaling_scale"),
+                    "skip_block_list": first_pass_args.get("skip_block_list"),
+                })
+                schedule = first_pass_args.get("timesteps") or first_pass_args.get("guidance_timesteps")
+                if schedule:
+                    first_pass_kwargs["timesteps"] = schedule
+                    first_pass_kwargs["guidance_timesteps"] = schedule
+                
+                downscale_factor = self.config.get("downscale_factor", 2)
+                original_height = first_pass_kwargs["height"]
+                original_width = first_pass_kwargs["width"]
+                divisor = 24
+
+                target_height_p1 = original_height // downscale_factor
+                height_p1 = round(target_height_p1 / divisor) * divisor
+                if height_p1 == 0: height_p1 = divisor
+                first_pass_kwargs["height"] = height_p1
+
+                target_width_p1 = original_width // downscale_factor
+                width_p1 = round(target_width_p1 / divisor) * divisor
+                if width_p1 == 0: width_p1 = divisor
+                first_pass_kwargs["width"] = width_p1
+                
+                print(f"[DEBUG] Passo 1: Dimensões reduzidas e ajustadas para {height_p1}x{width_p1}")
+                
                 ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
                 with ctx:
-                    result = multi_scale_pipeline(**multi_scale_call_kwargs)
-                print(f"[DEBUG] multi_scale_pipeline tempo={time.perf_counter()-t_ms:.3f}s")
+                    first_pass_result = self.pipeline(**first_pass_kwargs)
+                    
+                latents_low_res = first_pass_result.latents if hasattr(first_pass_result, "latents") else first_pass_result
+                log_tensor_info(latents_low_res, "Latentes (Passo 1)")
+                
+                del first_pass_result
+                gc.collect()
+                if self.device == "cuda": torch.cuda.empty_cache()
 
-                if hasattr(result, "latents"):
-                    latents = result.latents
-                elif hasattr(result, "images") and isinstance(result.images, torch.Tensor):
-                    latents = result.images
-                else:
-                    latents = result
-                print(f"[DEBUG] Latentes (multi-escala): shape={tuple(latents.shape)}")
+                # 2. Upscale dos latentes
+                print("[DEBUG] Multi-escala: Fazendo upscale dos latentes com latent_upsampler.")
+                with ctx:
+                    # Chamada posicional confirmada pelo código-fonte
+                    latents_high_res = self.latent_upsampler(latents_low_res)
+                
+                log_tensor_info(latents_high_res, "Latentes (Pós-Upscale)")
+                del latents_low_res
+                gc.collect()
+                if self.device == "cuda": torch.cuda.empty_cache()
+                
+                # 3. Configurar e executar o segundo passo
+                print("[DEBUG] Multi-escala: Iniciando Passo 2 (refinamento em alta resolução).")
+                second_pass_args = self.config.get("second_pass", {}).copy()
+                second_pass_kwargs = call_kwargs.copy()
+
+                # ==================== LÓGICA DE DIMENSÃO FINAL ====================
+                # As dimensões do Passo 2 DEVEM ser o dobro das dimensões do Passo 1,
+                # para corresponder à saída do upsampler.
+                height_p2 = height_p1 * 2
+                width_p2 = width_p1 * 2
+                second_pass_kwargs["height"] = height_p2
+                second_pass_kwargs["width"] = width_p2
+                print(f"[DEBUG] Passo 2: Dimensões definidas para {height_p2}x{width_p2} para corresponder ao upscale.")
+                # =================================================================
+
+                second_pass_kwargs.update({
+                    "guidance_scale": float(guidance_scale),
+                    "stg_scale": second_pass_args.get("stg_scale"),
+                    "rescaling_scale": second_pass_args.get("rescaling_scale"),
+                    "skip_block_list": second_pass_args.get("skip_block_list"),
+                })
+                
+                schedule_p2 = second_pass_args.get("timesteps") or second_pass_args.get("guidance_timesteps")
+                if schedule_p2:
+                    second_pass_kwargs["timesteps"] = schedule_p2
+                    second_pass_kwargs["guidance_timesteps"] = schedule_p2
+                    
+                second_pass_kwargs["latents"] = latents_high_res
+                
+                with ctx:
+                    second_pass_result = self.pipeline(**second_pass_kwargs)
+                    
+                latents = second_pass_result.latents if hasattr(second_pass_result, "latents") else second_pass_result
+                log_tensor_info(latents, "Latentes Finais (Passo 2)")
+                
             else:
                 single_pass_kwargs = call_kwargs.copy()
                 first_pass_config = self.config.get("first_pass", {})