lora_inference.py

from diffusers import AutoPipelineForText2Image
import torch

# Load the base pipeline
base_model = "runwayml/stable-diffusion-v1-5"

# Define a function to create a pipeline with the appropriate LoRA weights
def create_pipeline_with_lora(view):
    pipeline = AutoPipelineForText2Image.from_pretrained(base_model, torch_dtype=torch.float16).to("cuda")
    if view == "front":
        pipeline.load_lora_weights("/content/out_front/checkpoint-100", weight_name="pytorch_lora_weights.safetensors")
    elif view == "back":
        pipeline.load_lora_weights("/content/out_back/checkpoint-100", weight_name="pytorch_lora_weights.safetensors")
    elif view == "side":
        pipeline.load_lora_weights("/content/out_side/checkpoint-100", weight_name="pytorch_lora_weights.safetensors")
    else:
        raise ValueError("Unsupported view: {}".format(view))
    return pipeline

# Define the prompt
prompt = "high quality traffic in street."

# Generate images for each view
views = ["front", "back", "side"]
for view in views:
    pipeline = create_pipeline_with_lora(view)
    view_prompt = f"<{view} view> {prompt}"
    image = pipeline(view_prompt).images[0]
    image.save(f"traffic_{view}.png")

print("Images for all views generated and saved successfully.")