Update api/ltx/ltx_aduc_pipeline.py

api/ltx/ltx_aduc_pipeline.py

@@ -17,12 +17,19 @@ import traceback
 import warnings
 from pathlib import Path
 from typing import Dict, List, Optional, Tuple
-
 import torch
 import yaml
 import numpy as np
 from einops import rearrange
 from huggingface_hub import hf_hub_download
+from transformers import T5EncoderModel, T5Tokenizer
+from safetensors import safe_open
+from managers.gpu_manager import gpu_manager
+from transformers import T5EncoderModel, T5Tokenizer
+from safetensors import safe_open
+from managers.gpu_manager import gpu_manager
+
+
 
 # ==============================================================================
 # --- INITIAL SETUP & CONFIGURATION ---
@@ -36,40 +43,27 @@ logging.basicConfig(level=logging.INFO, format='[%(levelname)s] %(message)s')
 DEPS_DIR = Path("/data")
 LTX_VIDEO_REPO_DIR = DEPS_DIR / "LTX-Video"
 BASE_CONFIG_PATH = LTX_VIDEO_REPO_DIR / "configs"
-DEFAULT_CONFIG_FILE = BASE_CONFIG_PATH / "ltxv-13b-0.9.8-distilled-fp8.yaml"
+DEFAULT_CONFIG_FILE = BASE_CONFIG_PATH / "ltxv-13b-0.9.8-dev-fp8.yaml"
 LTX_REPO_ID = "Lightricks/LTX-Video"
 RESULTS_DIR = Path("/app/output")
 DEFAULT_FPS = 24.0
 FRAMES_ALIGNMENT = 8
 
 # --- CRITICAL: DEPENDENCY PATH INJECTION ---
-def add_deps_to_path():
-    """Adds the LTX repository directory to the Python system path for imports."""
-    repo_path = str(LTX_VIDEO_REPO_DIR.resolve())
-    if repo_path not in sys.path:
-        sys.path.insert(0, repo_path)
-        logging.info(f"LTX-Video repository added to sys.path: {repo_path}")
-
-add_deps_to_path()
-
-# --- PROJECT IMPORTS ---
-try:
-    from ltx_video.pipelines.pipeline_ltx_video import LTXVideoPipeline # E outros...
-    from ltx_video.models.autoencoders.causal_video_autoencoder import CausalVideoAutoencoder
-    from ltx_video.models.transformers.transformer3d import Transformer3DModel
-    from ltx_video.models.transformers.symmetric_patchifier import SymmetricPatchifier
-    from ltx_video.schedulers.rf import RectifiedFlowScheduler
-    from transformers import T5EncoderModel, T5Tokenizer
-    from safetensors import safe_open
-    from managers.gpu_manager import gpu_manager
-    from ltx_video.models.autoencoders.vae_encode import (normalize_latents, un_normalize_latents)
-    from ltx_video.pipelines.pipeline_ltx_video import (ConditioningItem, LTXMultiScalePipeline, adain_filter_latent)
-    from ltx_video.utils.inference_utils import load_image_to_tensor_with_resize_and_crop
-    from managers.vae_manager import vae_manager_singleton
-    from tools.video_encode_tool import video_encode_tool_singleton
-except ImportError as e:
-    logging.critical(f"A crucial LTX import failed. Check LTX-Video repo integrity. Error: {e}")
-    sys.exit(1)
+
+repo_path = str(LTX_VIDEO_REPO_DIR.resolve())
+if repo_path not in sys.path:
+    sys.path.insert(0, repo_path)
+    logging.info(f"LTX-Video repository added to sys.path: {repo_path}")
+
+from ltx_video.pipelines.pipeline_ltx_video import LTXVideoPipeline # E outros...
+from ltx_video.models.autoencoders.causal_video_autoencoder import CausalVideoAutoencoder
+from ltx_video.models.transformers.transformer3d import Transformer3DModel
+from ltx_video.models.transformers.symmetric_patchifier import SymmetricPatchifier
+from ltx_video.schedulers.rf import RectifiedFlowScheduler
+from ltx_video.models.autoencoders.vae_encode import (normalize_latents, un_normalize_latents)
+from ltx_video.pipelines.pipeline_ltx_video import (ConditioningItem, LTXMultiScalePipeline, adain_filter_latent)
+
 
 # ==============================================================================
 # --- UTILITY & HELPER FUNCTIONS ---
@@ -637,9 +631,22 @@ class LtxAducPipeline:
             conditioning_items.append(ConditioningItem(tensor, safe_frame, float(weight)))
         return conditioning_items
 
+    def _resize_tensor(self, media_items, height, width):
+        n_frames = media_items.shape[2]
+        if media_items.shape[-2:] != (height, width):
+            media_items = rearrange(media_items, "b c n h w -> (b n) c h w")
+            media_items = F.interpolate(
+                media_items,
+                size=(height, width),
+                mode="bilinear",
+                align_corners=False,
+            )
+            media_items = rearrange(media_items, "(b n) c h w -> b c n h w", n=n_frames)
+        return media_items
+
     def _prepare_conditioning_tensor(self, media_path: str, height: int, width: int, padding: Tuple) -> torch.Tensor:
         """Loads and processes an image to be a conditioning tensor."""
-        tensor = load_image_to_tensor_with_resize_and_crop(media_path, height, width)
+        tensor = self._resize_tensor(media_path, height, width)
         tensor = torch.nn.functional.pad(tensor, padding)
         # Conditioning tensors are needed on the main device for the transformer pass
         return tensor.to(self.main_device, dtype=self.runtime_autocast_dtype)