Create README.md

README.md

@@ -0,0 +1,99 @@
+---
+base_model:
+- Lightricks/LTX-Video
+library_name: diffusers
+datasets:
+- nexar-ai/nexar_collision_prediction
+---
+
+
+# Traffic World Models (LTX-Video, i2v)
+
+This repository hosts the fine-tuned LTX-Video image-to-video (i2v) diffusion model specialized for generating realistic driving footage.
+
+## Model Details
+
+- **Base Model**: LTX-Video
+- **Resolution**: 768×512 pixels (Adjustable)
+- **Frame Length**: 49 frames per generated video (Adjustable)
+- **Fine-tuning Method**: Low-Rank Adaptation (LoRA)
+- **Data Source**: Nexar Collision Prediction dataset videos (1,500 clips)
+
+## Usage Example
+
+```python
+import os
+import argparse
+import torch
+from diffusers.utils import export_to_video, load_image
+from stg_ltx_i2v_pipeline import LTXImageToVideoSTGPipeline
+
+def generate_video_from_image(
+    image_path,
+    prompt,
+    output_dir="outputs",
+    width=768,
+    height=512,
+    num_frames=49,
+    lora_path="mehmetkeremturkcan/DashcamCrashModels-LTX-I2V",
+    lora_weight=1.0,
+    prefix="crashmodel, ",
+    negative_prompt="worst quality, inconsistent motion, blurry, jittery, distorted",
+    stg_mode="STG-A",
+    stg_applied_layers_idx=[19],
+    stg_scale=1.0,
+    do_rescaling=True
+):
+    # Create output directory if it doesn't exist
+    if not os.path.exists(output_dir):
+        os.makedirs(output_dir)
+    # Load the model
+    pipe = LTXImageToVideoSTGPipeline.from_pretrained(
+        "a-r-r-o-w/LTX-Video-0.9.1-diffusers", 
+        torch_dtype=torch.bfloat16, 
+        local_files_only=False
+    )
+    # Apply LoRA weights
+    pipe.load_lora_weights(
+        lora_path, 
+        weight_name="pytorch_lora_weights.safetensors", 
+        adapter_name="suturing"
+    )
+    pipe.set_adapters("suturing", lora_weight)
+    pipe.to("cuda")
+    # Prepare the image and prompt
+    image = load_image(image_path).resize((width, height))
+    full_prompt = prefix + prompt if prefix else prompt
+    # Generate output filename
+    basename = os.path.basename(image_path).split('.')[0]
+    output_filename = f"{basename}_i2v.mp4"
+    output_path = os.path.join(output_dir, output_filename)
+    # Generate the video
+    print(f"Generating video with prompt: {full_prompt}")
+    video = pipe(
+        image=image,
+        prompt=full_prompt,
+        negative_prompt=negative_prompt,
+        width=width,
+        height=height,
+        num_frames=num_frames,
+        num_inference_steps=50,
+        decode_timestep=0.03,
+        decode_noise_scale=0.025,
+        generator=None,
+        stg_mode=stg_mode,
+        stg_applied_layers_idx=stg_applied_layers_idx,
+        stg_scale=stg_scale,
+        do_rescaling=do_rescaling
+    ).frames[0]
+    
+    # Export the video
+    export_to_video(video, output_path, fps=24)
+    print(f"Video saved to: {output_path}")
+    return output_path
+
+generate_video_from_image(
+    image_path="../suturing_datasetv2/images/9_railroad_final_8487-8570_NeedleWithdrawalNonIdeal.png",
+    prompt="A needlewithdrawalnonideal clip, generated from a backhand task."
+)
+```