mehmetkeremturkcan/DashcamCrashModels-LTX-I2V

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, 121 recommended)
• Fine-tuning Method: Low-Rank Adaptation (LoRA)
• Data Source: Nexar Collision Prediction dataset videos (1,500 clips)

Usage Example

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=121,
    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="crashmodel"
    )
    pipe.set_adapters("crashmodel", 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="my_starting_image.png",
    prompt="Dashcam footage of a near-miss in a Highway environment with Snow weather and Dark lighting conditions." # "near-miss" for crashes, "regular driving" for regular driving footage
)