Fixed Loading

app.py

@@ -1,16 +1,8 @@
 import spaces
 import torch
-from diffusers import AutoencoderKLWan, WanImageToVideoPipeline, UniPCMultistepScheduler, WanTransformer3DModel
+from diffusers import AutoencoderKLWan, WanImageToVideoPipeline, UniPCMultistepScheduler, WanTransformer3DModel, AutoModel, DiffusionPipeline
 from diffusers.utils import export_to_video
-# Conditionally import T2V pipeline to handle different diffusers versions and prevent crashes.
-try:
-    from diffusers import WanTextToVideoPipeline
-    IS_T2V_AVAILABLE = True
-except ImportError:
-    WanTextToVideoPipeline = None # Define as None so later code doesn't raise NameError
-    IS_T2V_AVAILABLE = False
-    print("⚠️ Warning: 'WanTextToVideoPipeline' could not be imported. Your 'diffusers' version might be outdated (requires >= 0.25.0).")
-from transformers import CLIPVisionModel, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPVisionModel, UMT5EncoderModel
 from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline # noqa
 import tempfile
 import re
@@ -24,24 +16,27 @@ import gradio as gr
 import random
 
 # --- I2V (Image-to-Video) Configuration ---
-I2V_MODEL_ID = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers" # Used for VAE/encoder components
-I2V_SINGLE_FILE_URL = "https://huggingface.co/vrgamedevgirl84/Wan14BT2VFusioniX/resolve/main/Wan14Bi2vFusioniX.safetensors"
+I2V_BASE_MODEL_ID = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers" # Used for VAE/encoder components
+I2V_FUSIONX_REPO_ID = "vrgamedevgirl84/Wan14BT2VFusioniX"
+I2V_FUSIONX_FILENAME = "Wan14Bi2vFusioniX.safetensors"
 
 # --- T2V (Text-to-Video) Configuration ---
-T2V_MODEL_ID = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
-T2V_SINGLE_FILE_URL = "https://huggingface.co/vrgamedevgirl84/Wan14BT2VFusioniX/resolve/main/WanT2V_MasterModel.safetensors"
+T2V_BASE_MODEL_ID = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
+T2V_LORA_REPO_ID = "vrgamedevgirl84/Wan14BT2VFusioniX"
+T2V_LORA_FILENAME = "FusionX_LoRa/Wan2.1_T2V_14B_FusionX_LoRA.safetensors"
 
 # --- Load Pipelines ---
 print("🚀 Loading I2V pipeline from single file...")
 i2v_pipe = None
 try:
     # Load components needed for the pipeline from the base model repo
-    i2v_image_encoder = CLIPVisionModel.from_pretrained(I2V_MODEL_ID, subfolder="image_encoder", torch_dtype=torch.float32)
-    i2v_vae = AutoencoderKLWan.from_pretrained(I2V_MODEL_ID, subfolder="vae", torch_dtype=torch.float32)
+    i2v_image_encoder = CLIPVisionModel.from_pretrained(I2V_BASE_MODEL_ID, subfolder="image_encoder", torch_dtype=torch.float32)
+    i2v_vae = AutoencoderKLWan.from_pretrained(I2V_BASE_MODEL_ID, subfolder="vae", torch_dtype=torch.float32)
     
-    # Load the main transformer from the single file URL
+    # Load the main transformer from the repo and filename
     i2v_transformer = WanTransformer3DModel.from_single_file(
-        I2V_SINGLE_FILE_URL,
+        I2V_FUSIONX_REPO_ID,
+        filename=I2V_FUSIONX_FILENAME,
         torch_dtype=torch.bfloat16
     )
 
@@ -58,34 +53,37 @@ except Exception as e:
     print(f"❌ Critical Error: Failed to load I2V pipeline from single file.")
     traceback.print_exc()
 
-print("\n🚀 Loading T2V pipeline from single file...")
+print("\n🚀 Loading T2V pipeline with LoRA...")
 t2v_pipe = None
-if IS_T2V_AVAILABLE:
-    try:
-        # Load components needed for the pipeline from the base model repo
-        t2v_vae = AutoencoderKLWan.from_pretrained(T2V_MODEL_ID, subfolder="vae", torch_dtype=torch.float32)
-        t2v_text_encoder = CLIPTextModel.from_pretrained(T2V_MODEL_ID, subfolder="text_encoder", torch_dtype=torch.bfloat16)
-        t2v_tokenizer = CLIPTokenizer.from_pretrained(T2V_MODEL_ID, subfolder="tokenizer")
-
-        # Load the main transformer from the single file URL
-        t2v_transformer = WanTransformer3DModel.from_single_file(
-            T2V_SINGLE_FILE_URL,
-            torch_dtype=torch.bfloat16
-        )
+try:
+    
+    # Load components needed for the T2V pipeline    
+    text_encoder = UMT5EncoderModel.from_pretrained(T2V_BASE_MODEL_ID, subfolder="text_encoder", torch_dtype=torch.bfloat16)
+    vae = AutoModel.from_pretrained(T2V_BASE_MODEL_ID, subfolder="vae", torch_dtype=torch.float32)
+    transformer = AutoModel.from_pretrained(T2V_BASE_MODEL_ID, subfolder="transformer", torch_dtype=torch.bfloat16)
+    
+    # Assemble the final pipeline
+    t2v_pipe = DiffusionPipeline.from_pretrained(
+        "Wan-AI/Wan2.1-T2V-14B-Diffusers",
+        vae=vae,
+        transformer=transformer,
+        text_encoder=text_encoder,
+        torch_dtype=torch.bfloat16
+    )
+    t2v_pipe.to("cuda")
 
-        # Manually assemble the pipeline with the custom transformer
-        t2v_pipe = WanTextToVideoPipeline(
-            vae=t2v_vae,
-            text_encoder=t2v_text_encoder,
-            tokenizer=t2v_tokenizer,
-            transformer=t2v_transformer
-        )
-        t2v_pipe.scheduler = UniPCMultistepScheduler.from_config(t2v_pipe.scheduler.config, flow_shift=8.0)
-        t2v_pipe.to("cuda")
-        print("✅ T2V pipeline loaded successfully from single file.")
-    except Exception as e:
-        print(f"❌ Critical Error: Failed to load T2V pipeline from single file.")
-        traceback.print_exc()
+    t2v_pipe.load_lora_weights(
+        T2V_LORA_REPO_ID,
+        weight_name=T2V_LORA_FILENAME,
+        adapter_name="fusionx_t2v"
+    )
+    t2v_pipe.set_adapters(["fusionx_t2v"], adapter_weights=[0.75])
+
+
+    print("✅ T2V pipeline and LoRA loaded and fused successfully.")
+except Exception as e:
+    print(f"❌ Critical Error: Failed to load T2V pipeline.")
+    traceback.print_exc()
 
 # --- LLM Prompt Enhancer Setup ---
 print("\n🤖 Loading LLM for Prompt Enhancement (Qwen/Qwen3-8B)...")
@@ -149,6 +147,7 @@ SLIDER_MIN_W, SLIDER_MAX_W = 128, 1024
 MAX_SEED = np.iinfo(np.int32).max
 
 FIXED_FPS = 24
+T2V_FIXED_FPS = 16
 MIN_FRAMES_MODEL = 8
 MAX_FRAMES_MODEL = 81
 
@@ -462,7 +461,7 @@ def generate_t2v_video(prompt, height, width,
 
     target_h = max(MOD_VALUE, (int(height) // MOD_VALUE) * MOD_VALUE)
     target_w = max(MOD_VALUE, (int(width) // MOD_VALUE) * MOD_VALUE)
-    num_frames = np.clip(int(round(duration_seconds * FIXED_FPS)), MIN_FRAMES_MODEL, MAX_FRAMES_MODEL)
+    num_frames = np.clip(int(round(duration_seconds * T2V_FIXED_FPS)), MIN_FRAMES_MODEL, MAX_FRAMES_MODEL)
     current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
     enhanced_prompt = f"{prompt}, cinematic, high detail, professional lighting"
 
@@ -482,7 +481,7 @@ def generate_t2v_video(prompt, height, width,
     filename = f"t2v_{sanitized_prompt}_{current_seed}.mp4"
     temp_dir = tempfile.mkdtemp()
     video_path = os.path.join(temp_dir, filename)
-    export_to_video(output_frames_list, video_path, fps=FIXED_FPS)
+    export_to_video(output_frames_list, video_path, fps=T2V_FIXED_FPS)
     
     return video_path, current_seed, gr.File(value=video_path, visible=True, label=f"📥 Download: {filename}")
 
@@ -530,7 +529,7 @@ with gr.Blocks(css=custom_css) as demo:
 
             # --- Text-to-Video Tab ---
             with gr.TabItem("✍️ Text-to-Video", id="t2v_tab", interactive=t2v_pipe is not None):
-                if not IS_T2V_AVAILABLE or t2v_pipe is None:
+                if t2v_pipe is None:
                     gr.Markdown("<h3 style='color: #ff9999; text-align: center;'>⚠️ Text-to-Video Pipeline Failed to Load. This tab is disabled.</h3>")
                 else:
                     with gr.Row():
@@ -548,7 +547,7 @@ with gr.Blocks(css=custom_css) as demo:
                                 minimum=round(MIN_FRAMES_MODEL/FIXED_FPS,1),
                                 maximum=round(MAX_FRAMES_MODEL/FIXED_FPS,1),
                                 step=0.1, value=2, label="⏱️ Duration (seconds)",
-                                info=f"Generates {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps."
+                                info=f"Generates {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {T2V_FIXED_FPS}fps."
                             )
                             with gr.Accordion("⚙️ Advanced Settings", open=False):
                                 t2v_neg_prompt = gr.Textbox(label="❌ Negative Prompt", value=default_negative_prompt, lines=4)
@@ -558,7 +557,7 @@ with gr.Blocks(css=custom_css) as demo:
                                     t2v_height = gr.Slider(minimum=SLIDER_MIN_H, maximum=SLIDER_MAX_H, step=MOD_VALUE, value=DEFAULT_H_SLIDER_VALUE, label=f"📏 Height ({MOD_VALUE}px steps)")
                                     t2v_width = gr.Slider(minimum=SLIDER_MIN_W, maximum=SLIDER_MAX_W, step=MOD_VALUE, value=DEFAULT_W_SLIDER_VALUE, label=f"📐 Width ({MOD_VALUE}px steps)")
                                 t2v_steps = gr.Slider(minimum=1, maximum=25, step=1, value=15, label="🚀 Inference Steps", info="15-20 recommended for quality.")
-                                t2v_guidance = gr.Slider(minimum=0.0, maximum=20.0, step=0.5, value=7.5, label="🎯 Guidance Scale")
+                                t2v_guidance = gr.Slider(minimum=0.0, maximum=20.0, step=0.5, value=5.0, label="🎯 Guidance Scale")
                             
                             t2v_generate_btn = gr.Button("🎬 Generate T2V", variant="primary", elem_classes=["generate-btn"])