Awesome new Lora

app.py

@@ -5,6 +5,8 @@ from diffusers.utils import export_to_video
 from transformers import CLIPVisionModel
 import gradio as gr
 import tempfile
+import re
+import os
 
 from huggingface_hub import hf_hub_download
 import numpy as np
@@ -12,15 +14,12 @@ from PIL import Image
 import random
 
 # Base MODEL_ID (using original Wan model that's compatible with diffusers)
-MODEL_ID = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
+MODEL_ID = "Wan-AI/Wan2.1-I2V-14B-720P-Diffusers"
 
-# FusionX enhancement LoRAs (based on FusionX composition)
-LORA_REPO_ID = "Kijai/WanVideo_comfy"
-LORA_FILENAME = "Wan21_CausVid_14B_T2V_lora_rank32.safetensors"
-
-# Additional enhancement LoRAs for FusionX-like quality
-ACCVIDEO_LORA_REPO = "alibaba-pai/Wan2.1-Fun-Reward-LoRAs"
-MPS_LORA_FILENAME = "Wan2.1-Fun-14B-InP-MPS.safetensors"
+# Merged FusionX enhancement LoRA
+LORA_REPO_ID = "vrgamedevgirl84/Wan14BT2VFusioniX"
+LORA_FILENAME = "Wan2.1_I2V_14B_FusionX_LoRA.safetensors"
+LORA_SUBFOLDER = "FusionX_LoRa"
 
 # Load enhanced model components
 print("🚀 Loading FusionX Enhanced Wan2.1 I2V Model...")
@@ -34,37 +33,18 @@ pipe = WanImageToVideoPipeline.from_pretrained(
 pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=8.0)
 pipe.to("cuda")
 
-# Load FusionX enhancement LoRAs
-lora_adapters = []
-lora_weights = []
-
-try:
-    # Load CausVid LoRA (strength 1.0 as per FusionX)
-    causvid_path = hf_hub_download(repo_id=LORA_REPO_ID, filename=LORA_FILENAME)
-    pipe.load_lora_weights(causvid_path, adapter_name="causvid_lora")
-    lora_adapters.append("causvid_lora")
-    lora_weights.append(1.0)  # FusionX uses 1.0 for CausVid
-    print("✅ CausVid LoRA loaded (strength: 1.0)")
-except Exception as e:
-    print(f"⚠️ CausVid LoRA not loaded: {e}")
-
+# Load and fuse the single merged FusionX LoRA
 try:
-    # Load MPS Rewards LoRA (strength 0.7 as per FusionX)
-    mps_path = hf_hub_download(repo_id=ACCVIDEO_LORA_REPO, filename=MPS_LORA_FILENAME)
-    pipe.load_lora_weights(mps_path, adapter_name="mps_lora")
-    lora_adapters.append("mps_lora")
-    lora_weights.append(0.7)  # FusionX uses 0.7 for MPS
-    print("✅ MPS Rewards LoRA loaded (strength: 0.7)")
+    lora_path = hf_hub_download(
+        repo_id=LORA_REPO_ID, 
+        filename=LORA_FILENAME, 
+        subfolder=LORA_SUBFOLDER
+    )
+    pipe.load_lora_weights(lora_path, adapter_name="fusionx")
+    print("✅ Merged FusionX LoRA loaded. Use the 'LoRA Strength' slider to control the effect.")
 except Exception as e:
-    print(f"⚠️ MPS LoRA not loaded: {e}")
-
-# Apply LoRA adapters if any were loaded
-if lora_adapters:
-    pipe.set_adapters(lora_adapters, adapter_weights=lora_weights)
-    pipe.fuse_lora()
-    print(f"🔥 FusionX Enhancement Applied: {len(lora_adapters)} LoRAs fused")
-else:
-    print("📝 No LoRAs loaded - using base Wan model")
+    print(f"⚠️ Merged FusionX LoRA not loaded: {e}")
+    print("📝 Using base Wan model without LoRA enhancement. The 'LoRA Strength' slider will have no effect.")
 
 MOD_VALUE = 32
 DEFAULT_H_SLIDER_VALUE = 576  # FusionX optimized default
@@ -288,6 +268,17 @@ input[type="checkbox"] {
 }
 """
 
+def sanitize_prompt_for_filename(prompt: str, max_len: int = 60) -> str:
+    """Sanitizes a prompt string to be used as a valid filename."""
+    if not prompt:
+        prompt = "video"
+    # Remove non-alphanumeric characters (except spaces, hyphens, underscores)
+    sanitized = re.sub(r'[^\w\s_-]', '', prompt).strip()
+    # Replace spaces and multiple hyphens/underscores with a single underscore
+    sanitized = re.sub(r'[\s_-]+', '_', sanitized)
+    # Truncate to max_len
+    return sanitized[:max_len]
+
 def _calculate_new_dimensions_wan(pil_image, mod_val, calculation_max_area,
                                  min_slider_h, max_slider_h,
                                  min_slider_w, max_slider_w,
@@ -325,7 +316,7 @@ def handle_image_upload_for_dims_wan(uploaded_pil_image, current_h_val, current_
 
 def get_duration(input_image, prompt, height, width, 
                    negative_prompt, duration_seconds,
-                   guidance_scale, steps,
+                   guidance_scale, steps, lora_scale,
                    seed, randomize_seed, 
                    progress):
     # FusionX optimized duration calculation
@@ -339,7 +330,7 @@ def get_duration(input_image, prompt, height, width,
 @spaces.GPU(duration=get_duration)
 def generate_video(input_image, prompt, height, width, 
                    negative_prompt=default_negative_prompt, duration_seconds=3,
-                   guidance_scale=1, steps=8,  # FusionX optimized default
+                   guidance_scale=1, steps=8, lora_scale=1.0,
                    seed=42, randomize_seed=False, 
                    progress=gr.Progress(track_tqdm=True)):
     
@@ -368,11 +359,17 @@ def generate_video(input_image, prompt, height, width,
             num_frames=num_frames,
             guidance_scale=float(guidance_scale), 
             num_inference_steps=int(steps),
-            generator=torch.Generator(device="cuda").manual_seed(current_seed)
+            generator=torch.Generator(device="cuda").manual_seed(current_seed),
+            cross_attention_kwargs={"scale": float(lora_scale)}
         ).frames[0]
 
-    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
-        video_path = tmpfile.name
+    # Create a unique filename for download
+    sanitized_prompt = sanitize_prompt_for_filename(prompt)
+    filename = f"{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)
     return video_path, current_seed
 
@@ -439,6 +436,14 @@ with gr.Blocks() as demo:
                             value=DEFAULT_W_SLIDER_VALUE, 
                             label=f"📐 Output Width (FusionX optimized: {MOD_VALUE} multiples)"
                         )
+                    lora_scale_slider = gr.Slider(
+                        minimum=0.0,
+                        maximum=2.5,
+                        step=0.05,
+                        value=1.0,
+                        label="💪 FusionX LoRA Strength",
+                        info="Control the intensity of the FusionX effect. >1.0 for stronger effect, <1.0 for less."
+                    )
                     steps_slider = gr.Slider(
                         minimum=1, 
                         maximum=20, 
@@ -466,7 +471,8 @@ with gr.Blocks() as demo:
                 video_output = gr.Video(
                     label="🎥 FusionX Enhanced Generated Video", 
                     autoplay=True, 
-                    interactive=False
+                    interactive=False,
+                    download=True
                 )
 
         input_image_component.upload(
@@ -484,23 +490,10 @@ with gr.Blocks() as demo:
         ui_inputs = [
             input_image_component, prompt_input, height_input, width_input,
             negative_prompt_input, duration_seconds_input,
-            guidance_scale_input, steps_slider, seed_input, randomize_seed_checkbox
+            guidance_scale_input, steps_slider, lora_scale_slider, seed_input, randomize_seed_checkbox
         ]
         generate_button.click(fn=generate_video, inputs=ui_inputs, outputs=[video_output, seed_input])
 
-        with gr.Column():
-            gr.Examples(
-                examples=[ 
-                    ["peng.png", "a penguin gracefully dancing in the pristine snow, cinematic motion with detailed feathers", 576, 576],
-                    ["frog.jpg", "the frog jumps energetically with smooth, lifelike motion and detailed texture", 576, 576],
-                ],
-                inputs=[input_image_component, prompt_input, height_input, width_input], 
-                outputs=[video_output, seed_input], 
-                fn=generate_video, 
-                cache_examples="lazy",
-                label="🌟 FusionX Enhanced Example Gallery"
-            )
-        
 
 if __name__ == "__main__":
     demo.queue().launch()