app.py

import spaces
import os
import time
import torch
import gradio as gr
from PIL import Image
from huggingface_hub import hf_hub_download, list_repo_files, login
from src_inference.pipeline import FluxPipeline
from src_inference.lora_helper import set_single_lora

HF_TOKEN = os.environ.get("HF_TOKEN") 
if HF_TOKEN:
    login(token=HF_TOKEN)

BASE_PATH = "black-forest-labs/FLUX.1-dev"
LOCAL_LORA_DIR = "./LoRAs"          
CUSTOM_LORA_DIR = "./Custom_LoRAs"  
os.makedirs(LOCAL_LORA_DIR, exist_ok=True)
os.makedirs(CUSTOM_LORA_DIR, exist_ok=True)

print("downloading OmniConsistency base LoRA …")
omni_consistency_path = hf_hub_download(
    repo_id="showlab/OmniConsistency",
    filename="OmniConsistency.safetensors",
    local_dir="./Model"
)

print("loading base pipeline …")
pipe = FluxPipeline.from_pretrained(
    BASE_PATH, torch_dtype=torch.bfloat16
).to("cuda")
set_single_lora(pipe.transformer, omni_consistency_path,
                lora_weights=[1], cond_size=512)

lora_names = [
    "3D_Chibi", "American_Cartoon", "Macaron",
    "Pixel", "Poly", "Van_Gogh"
]

def download_all_loras():
    for name in lora_names:
        hf_hub_download(
            repo_id="showlab/OmniConsistency",
            filename=f"LoRAs/{name}_rank128_bf16.safetensors",
            local_dir=LOCAL_LORA_DIR,
        )
download_all_loras()

def reload_all_loras():
    pipe.unload_lora_weights()
    for name in lora_names:
        pipe.load_lora_weights(
            f"{LOCAL_LORA_DIR}/LoRAs",
            weight_name=f"{name}_rank128_bf16.safetensors",
            adapter_name=name,
        )
reload_all_loras()
        
def clear_cache(transformer):
    for _, attn_processor in transformer.attn_processors.items():
        attn_processor.bank_kv.clear()

@spaces.GPU()
def generate_image(
    lora_name,
    prompt,
    uploaded_image,
    width, height,
    guidance_scale,
    num_inference_steps,
    seed
):
    width, height = int(width), int(height)
    generator = torch.Generator("cpu").manual_seed(seed) 

    pipe.set_adapters(lora_name)

    spatial_image  = [uploaded_image.convert("RGB")]
    subject_images = []
    start = time.time()
    out_img = pipe(
        prompt,
        height=(height // 8) * 8,
        width=(width  // 8) * 8,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        max_sequence_length=512,
        generator=generator,
        spatial_images=spatial_image,
        subject_images=subject_images,
        cond_size=512,
    ).images[0]
    print(f"inference time: {time.time()-start:.2f}s")

    clear_cache(pipe.transformer)
    return uploaded_image, out_img

# =============== Gradio UI ===============
def create_interface():

    def update_trigger_word(lora_name, prompt):
      for name in lora_names:
        trigger = " ".join(name.split("_")) + " style,"
        prompt = prompt.replace(trigger, "")
      new_trigger = " ".join(lora_name.split("_"))+ " style,"
      return new_trigger + prompt


    header = """
<div style="text-align: center; display: flex; justify-content: left; gap: 5px;">
<a href="https://arxiv.org/abs/2505.18445"><img src="https://img.shields.io/badge/ariXv-2505.18445-A42C25.svg" alt="arXiv"></a>
<a href="https://huggingface.co/showlab/OmniConsistency"><img src="https://img.shields.io/badge/🤗_HuggingFace-Model-ffbd45.svg" alt="HuggingFace"></a>
<a href="https://github.com/showlab/OmniConsistency"><img src="https://img.shields.io/badge/GitHub-Code-blue.svg?logo=github&" alt="GitHub"></a>
</div>
"""

    with gr.Blocks() as demo:
        gr.Markdown("# OmniConsistency LoRA Image Generation")
        gr.Markdown("Select a LoRA, enter a prompt, and upload an image to generate a new image with OmniConsistency.")
        gr.HTML(header)

        with gr.Row():
            with gr.Column(scale=1):
                image_input = gr.Image(type="pil", label="Upload Image")
                prompt_box = gr.Textbox(label="Prompt",
                    value="3D Chibi style,",
                    info="Remember to include the necessary trigger words if you're using a custom LoRA."
                )
                lora_dropdown = gr.Dropdown(
                    demo_lora_names, label="Select built-in LoRA")
                custom_repo_box = gr.Textbox(
                    label="Enter Custom LoRA repo",
                    placeholder="LoRA Hugging Face path (e.g., 'username/repo_name')",
                    info="If you want to use a custom LoRA, enter its Hugging Face repo ID here and built-in LoRA will be Overridden. Leave empty to use built-in LoRAs. [Check the list of FLUX LoRAs](https://huggingface.co/models?other=base_model:adapter:black-forest-labs/FLUX.1-dev)"
                )
                custom_weight_name = gr.Textbox(label="Enter Custom LoRA weight name",)
                gen_btn = gr.Button("Generate")
            with gr.Column(scale=1):
                output_image = gr.ImageSlider(label="Generated Image")
        with gr.Accordion("Advanced Options", open=False):
          height_box = gr.Textbox(value="1024", label="Height")
          width_box  = gr.Textbox(value="1024", label="Width")
          guidance_slider = gr.Slider(
              0.1, 20, value=3.5, step=0.1, label="Guidance Scale")
          steps_slider = gr.Slider(
              1, 50, value=25, step=1, label="Inference Steps")
          seed_slider = gr.Slider(
              1, 2_147_483_647, value=42, step=1, label="Seed")

        lora_dropdown.select(fn=update_trigger_word, inputs=[lora_dropdown,prompt_box], 
                             outputs=prompt_box)

        gen_btn.click(
            fn=generate_image,
            inputs=[lora_dropdown, custom_repo_box, custom_weight_name, prompt_box, image_input,
                    width_box, height_box, guidance_slider, steps_slider, seed_slider],
            outputs=output_image
        )
    return demo

if __name__ == "__main__":
    demo = create_interface()
    demo.launch(ssr_mode=False)