Update api/ltx/ltx_aduc_pipeline.py

api/ltx/ltx_aduc_pipeline.py

@@ -150,8 +150,7 @@ class LtxAducPipeline:
             elif stg_mode_str.lower() in ["stg_r", "residual"]: stg_strategy = SkipLayerStrategy.Residual
             elif stg_mode_str.lower() in ["stg_t", "transformer_block"]: stg_strategy = SkipLayerStrategy.TransformerBlock
 
-            
-            
+        
             height_padded = ((kwargs['height'] - 1) // 8 + 1) * 8
             width_padded = ((kwargs['width'] - 1) // 8 + 1) * 8
             downscale_factor = self.config.get("downscale_factor", 0.6666666)
@@ -161,21 +160,37 @@ class LtxAducPipeline:
             x_height = int(height_padded * downscale_factor)
             downscaled_height = x_height - (x_height % vae_scale_factor)
             
-            
             call_kwargs = {
                 "height": downscaled_height, 
                 "width": downscaled_width,
-                "skip_initial_inference_steps": 0, "skip_final_inference_steps": 0, "num_inference_steps": 20,
+                "skip_initial_inference_steps": 3, 
+                "skip_final_inference_steps": 0, 
+                "num_inference_steps": 30,
                 "negative_prompt": kwargs['negative_prompt'], 
-                "guidance_scale": 4, "stg_scale": self.config.get("stg_scale", 4),
-                "rescaling_scale": self.config.get("rescaling_scale", 0.7), "skip_layer_strategy": stg_strategy,
-                "skip_block_list": self.config.get("skip_block_list", None), "frame_rate": int(DEFAULT_FPS),
+                "guidance_scale": self.config.get("guidance_scale", [1, 1, 6, 8, 6, 1, 1]), 
+                "stg_scale": self.config.get("stg_scale", [0, 0, 4, 4, 4, 2, 1]),
+                "rescaling_scale": self.config.get("rescaling_scale", [1, 1, 0.5, 0.5, 1, 1, 1]), 
+                "skip_block_list": self.config.get("skip_block_list", [[], [11, 25, 35, 39], [22, 35, 39], [28], [28], [28], [28]]), 
+                "frame_rate": int(DEFAULT_FPS),
                 "generator": torch.Generator(device=self.main_device).manual_seed(self._get_random_seed()),
-                "output_type": "latent", "media_items": None, "decode_timestep": self.config.get("decode_timestep", None),
-                "decode_noise_scale": self.config.get("decode_noise_scale", None), "stochastic_sampling": self.config.get("stochastic_sampling", None),
-                "image_cond_noise_scale": 0.15, "is_video": True, "vae_per_channel_normalize": True,
-                "mixed_precision": (self.config["precision"] == "mixed_precision"), "offload_to_cpu": False,
+                "output_type": "latent", 
+                "media_items": None, 
+                "decode_timestep": self.config.get("decode_timestep", 0.05),
+                "decode_noise_scale": self.config.get("decode_noise_scale", 0.025), 
+                "stochastic_sampling": self.config.get("stochastic_sampling", false),
+                "is_video": True, 
+                "vae_per_channel_normalize": True,
+                "offload_to_cpu": False,
                 "enhance_prompt": False,
+                "num_frames": total_frames,
+                "downscale_factor": self.config.get("downscale_factor", 0.6666666),
+                "rescaling_scale": self.config.get("rescaling_scale",  [1, 1, 0.5, 0.5, 1, 1, 1]),
+                "guidance_timesteps": self.config.get("guidance_timesteps", [1.0, 0.996,  0.9933, 0.9850, 0.9767, 0.9008, 0.6180]),
+                "skip_block_list": self.config.get("skip_block_list",  [[], [11, 25, 35, 39], [22, 35, 39], [28], [28], [28], [28]]),
+                "sampler": self.config.get("sampler", "from_checkpoint"),
+                "precision": self.config.get("precision", "float8_e4m3fn"),
+                "stochastic_sampling": self.config.get("stochastic_sampling", False),
+                "cfg_star_rescale": self.config.get("cfg_star_rescale", True),
             }
             
             ltx_configs_override = kwargs.get("ltx_configs_override", {})
@@ -185,11 +200,7 @@ class LtxAducPipeline:
             # --- ETAPA 1: GERAÇÃO DE CHUNKS E SALVAMENTO ---
             for i, chunk_prompt in enumerate(prompt_list):
                 logging.info(f"Processing scene {i+1}/{num_chunks}: '{chunk_prompt[:50]}...'")
-                current_frames_base = frames_per_chunk if i < num_chunks - 1 else total_frames - ((num_chunks - 1) * frames_per_chunk)
-                current_frames = current_frames_base + (overlap_frames if i > 0 else 0)
-                current_frames = self._align(current_frames, alignment_rule='n*8+1')
                 call_kwargs["prompt"] = chunk_prompt
-                call_kwargs["num_frames"] = current_frames
                 
                 with torch.autocast(device_type=self.main_device.type, dtype=self.runtime_autocast_dtype, enabled="cuda" in self.main_device.type):
                     chunk_latents = self.pipeline(**call_kwargs).images