app.py

import gradio as gr
from time import sleep, time
from diffusers import DiffusionPipeline, StableDiffusionXLPipeline
from huggingface_hub import hf_hub_download, CommitScheduler
from safetensors.torch import load_file
from share_btn import community_icon_html, loading_icon_html, share_js
from uuid import uuid4
from pathlib import Path
from PIL import Image
import torch
import json
import random
import copy
import gc
import pickle
import spaces

lora_list = hf_hub_download(repo_id="multimodalart/LoraTheExplorer", filename="sdxl_loras.json", repo_type="space")

IMAGE_DATASET_DIR = Path("image_dataset") / f"train-{uuid4()}"
IMAGE_DATASET_DIR.mkdir(parents=True, exist_ok=True)
IMAGE_JSONL_PATH = IMAGE_DATASET_DIR / "metadata.jsonl"

scheduler = CommitScheduler(
    repo_id="multimodalart/lora-fusing-preferences",
    repo_type="dataset",
    folder_path=IMAGE_DATASET_DIR,
    path_in_repo=IMAGE_DATASET_DIR.name,
    every=10
)

with open(lora_list, "r") as file:
    data = json.load(file)
    sdxl_loras = [
        {
            "image": item["image"] if item["image"].startswith("https://") else f'https://huggingface.co/spaces/multimodalart/LoraTheExplorer/resolve/main/{item["image"]}',
            "title": item["title"],
            "repo": item["repo"],
            "trigger_word": item["trigger_word"],
            "weights": item["weights"],
            "is_compatible": item["is_compatible"],
            "is_pivotal": item.get("is_pivotal", False),
            "text_embedding_weights": item.get("text_embedding_weights", None),
            "is_nc": item.get("is_nc", False)
        }
        for item in data
    ]

state_dicts = {}

for item in sdxl_loras:
    saved_name = hf_hub_download(item["repo"], item["weights"])
    
    if not saved_name.endswith('.safetensors'):
        state_dict = torch.load(saved_name, map_location=torch.device('cpu'))
    else:
        state_dict = load_file(saved_name, device="cpu")
    
    state_dicts[item["repo"]] = {
        "saved_name": saved_name,
        "state_dict": state_dict
    }

css = '''
.gradio-container{max-width: 650px! important}
#title{text-align:center;}
#title h1{font-size: 250%}
.selected_random img{object-fit: cover}
.selected_random [data-testid="block-label"] span{display: none}
.plus_column{align-self: center}
.plus_button{font-size: 235% !important; text-align: center;margin-bottom: 19px}
#prompt{padding: 1em}
#prompt input{width: calc(100% - 160px);border-top-right-radius: 0px;border-bottom-right-radius: 0px;}
#run_button{position: absolute;margin-top: 39.8px;right: 0;margin-right: 1em;border-bottom-left-radius: 0px;border-top-left-radius: 0px}
.random_column{align-self: center; align-items: center;gap: 0.5em !important}
#share-btn-container{padding-left: 0.5rem !important; padding-right: 0.5rem !important; background-color: #000000; justify-content: center; align-items: center; border-radius: 9999px !important; max-width: 13rem; margin-left: auto;margin-top: 0.35em;}
div#share-btn-container > div {flex-direction: row;background: black;align-items: center}
#share-btn-container:hover {background-color: #060606}
#share-btn {all: initial; color: #ffffff;font-weight: 600; cursor:pointer; font-family: 'IBM Plex Sans', sans-serif; margin-left: 0.5rem !important; padding-top: 0.5rem !important; padding-bottom: 0.5rem !important;right:0;font-size: 15px;}
#share-btn * {all: unset}
#share-btn-container div:nth-child(-n+2){width: auto !important;min-height: 0px !important;}
#share-btn-container .wrap {display: none !important}
#share-btn-container.hidden {display: none!important}
#post_gen_info{margin-top: .5em}
#thumbs_up_clicked{background:green}
#thumbs_down_clicked{background:red}
.title_lora a{color: var(--body-text-color) !important; opacity:0.6}
#prompt_area .form{border:0}
#reroll_button{position: absolute;right: 0;flex-grow: 1;min-width: 75px;padding: .1em}
.pending .min {min-height: auto}
'''

original_pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16)

@spaces.GPU
def merge_and_run(prompt, negative_prompt, shuffled_items, lora_1_scale=0.5, lora_2_scale=0.5, seed=-1, progress=gr.Progress(track_tqdm=True)):
  
  repo_id_1 = shuffled_items[0]['repo']
  repo_id_2 = shuffled_items[1]['repo']
  print("Loading state dicts...")
  start_time = time()  
  state_dict_1 = copy.deepcopy(state_dicts[repo_id_1]["state_dict"])
  state_dict_1 = {k: v.to(device="cuda", dtype=torch.float16) for k,v in state_dict_1.items() if torch.is_tensor(v)}
  state_dict_2 = copy.deepcopy(state_dicts[repo_id_2]["state_dict"])
  state_dict_2 = {k: v.to(device="cuda", dtype=torch.float16) for k,v in state_dict_2.items() if torch.is_tensor(v)}
  state_dict_time = time() - start_time
  print(f"State Dict time: {state_dict_time}")  
  start_time = time()
  unet = copy.deepcopy(original_pipe.unet)
  text_encoder=copy.deepcopy(original_pipe.text_encoder)
  text_encoder_2=copy.deepcopy(original_pipe.text_encoder_2)
  pipe = StableDiffusionXLPipeline(vae=original_pipe.vae,
                                   text_encoder=text_encoder,
                                   text_encoder_2=text_encoder_2,
                                   scheduler=original_pipe.scheduler,
                                   tokenizer=original_pipe.tokenizer,
                                   tokenizer_2=original_pipe.tokenizer_2,
                                   unet=unet)
  pickle_time = time() - start_time
  print(f"copy time: {pickle_time}")  
  pipe.to("cuda") 
  start_time = time()
  print("Loading LoRA weights...")
  pipe.load_lora_weights(state_dict_1, low_cpu_mem_usage=True)
  pipe.fuse_lora(lora_1_scale)
  pipe.load_lora_weights(state_dict_2, low_cpu_mem_usage=True)
  pipe.fuse_lora(lora_2_scale)
  lora_time = time() - start_time
  print(f"Loaded LoRAs time: {lora_time}") 
  if negative_prompt == "":
    negative_prompt = None
          
  if(seed < 0):
      seed = random.randint(0, 2147483647)
  generator = torch.Generator(device="cuda").manual_seed(seed)
  image = pipe(prompt=prompt, negative_prompt=negative_prompt, num_inference_steps=20, width=768, height=768, generator=generator).images[0]
  return image, gr.update(visible=True), seed, gr.update(visible=True, interactive=True), gr.update(visible=False), gr.update(visible=True, interactive=True), gr.update(visible=False)

def get_description(item):
      trigger_word = item["trigger_word"]
      return f"Trigger: `{trigger_word}`" if trigger_word else "No trigger, applied automatically", trigger_word

def truncate_string(s, max_length=29):
    return s[:max_length - 3] + "..." if len(s) > max_length else s

def shuffle_images():
    compatible_items = [item for item in sdxl_loras if item['is_compatible']]
    random.shuffle(compatible_items)
    two_shuffled_items = compatible_items[:2]
    print(two_shuffled_items)
    title_1 = gr.update(label=two_shuffled_items[0]['title'], value=two_shuffled_items[0]['image'])
    title_2 = gr.update(label=two_shuffled_items[1]['title'], value=two_shuffled_items[1]['image'])
    repo_id_1 = gr.update(value=two_shuffled_items[0]['repo'])
    repo_id_2 = gr.update(value=two_shuffled_items[1]['repo'])
    description_1, trigger_word_1 = get_description(two_shuffled_items[0])
    description_2, trigger_word_2 = get_description(two_shuffled_items[1])

    lora_1_link = f"[{truncate_string(two_shuffled_items[0]['repo'])}](https://huggingface.co/{two_shuffled_items[0]['repo']}) ✨"
    lora_2_link = f"[{truncate_string(two_shuffled_items[1]['repo'])}](https://huggingface.co/{two_shuffled_items[1]['repo']}) ✨"
    prompt_description_1 = gr.update(value=description_1, visible=True)
    prompt_description_2 = gr.update(value=description_2, visible=True)
    prompt = gr.update(value=f"{trigger_word_1} {trigger_word_2}")
    scale = gr.update(value=0.7)

    return lora_1_link, title_1, prompt_description_1, repo_id_1, lora_2_link, title_2, prompt_description_2, repo_id_2, prompt, two_shuffled_items, scale, scale

def save_preferences(lora_1_id, lora_1_scale, lora_2_id, lora_2_scale, prompt, generated_image, thumbs_direction, seed):
    image_path = IMAGE_DATASET_DIR / f"{uuid4()}.png"
    with scheduler.lock:
        Image.fromarray(generated_image).save(image_path)
        with IMAGE_JSONL_PATH.open("a") as f:
            json.dump({"prompt": prompt, "file_name":image_path.name, "lora_1_id": lora_1_id, "lora_1_scale": float(lora_1_scale), "lora_2_id": lora_2_id, "lora_2_scale": float(lora_2_scale), "thumbs_direction": thumbs_direction, "seed": int(seed)}, f)
            f.write("\n")
    
    return gr.update(visible=False), gr.update(visible=True), gr.update(interactive=False)

def hide_post_gen_info():
    return gr.update(visible=False)

with gr.Blocks(css=css) as demo:
  shuffled_items = gr.State()
  title = gr.HTML(
        '''<h1>LoRA Roulette 🎰</h1>
        <p>This random LoRAs are loaded into SDXL, can you find a fun way to combine them? 🎨</p>
        ''',
        elem_id="title"
  )
  with gr.Column():
    with gr.Column(min_width=10, scale=16, elem_classes="plus_column"):
        with gr.Row():
            with gr.Column(min_width=10, scale=4, elem_classes="random_column"):
              lora_1_link = gr.Markdown(elem_classes="title_lora")
              lora_1 = gr.Image(interactive=False, height=150, elem_classes="selected_random", elem_id="randomLoRA_1", show_share_button=False, show_download_button=False)
              lora_1_id = gr.Textbox(visible=False, elem_id="random_lora_1_id")
              lora_1_prompt = gr.Markdown(visible=False)
            with gr.Column(min_width=10, scale=1, elem_classes="plus_column"):
              plus = gr.HTML("+", elem_classes="plus_button")
            with gr.Column(min_width=10, scale=4, elem_classes="random_column"):
              lora_2_link = gr.Markdown(elem_classes="title_lora")
              lora_2 = gr.Image(interactive=False, height=150, elem_classes="selected_random", elem_id="randomLoRA_2", show_share_button=False, show_download_button=False)
              lora_2_id = gr.Textbox(visible=False, elem_id="random_lora_2_id")
              lora_2_prompt = gr.Markdown(visible=False)
            with gr.Column(min_width=10, scale=2, elem_classes="plus_column"):
               equal = gr.HTML("=", elem_classes="plus_button")
            shuffle_button = gr.Button("🎲  reroll", elem_id="reroll_button")
    with gr.Column(min_width=10, scale=14):
        with gr.Group(elem_id="generate_area"):
            with gr.Row(elem_id="prompt_area"):
                prompt = gr.Textbox(label="Your prompt", info="Rearrange the trigger words into a coherent prompt", show_label=False, interactive=True, elem_id="prompt")
                run_btn = gr.Button("Run", elem_id="run_button")
            output_image = gr.Image(label="Output", elem_id="output_image", interactive=False)
            with gr.Column(visible=False, elem_id="post_gen_info") as post_gen_info:
                with gr.Row():
                    with gr.Column(min_width=10):
                        thumbs_up = gr.Button("👍", elem_id="thumbs_up_unclicked")
                        thumbs_up_clicked = gr.Button("👍", elem_id="thumbs_up_clicked", interactive=False, visible=False)
                    with gr.Column(min_width=10):
                        thumbs_down = gr.Button("👎", elem_id="thumbs_down_unclicked")
                        thumbs_down_clicked = gr.Button("👎", elem_id="thumbs_down_clicked", interactive=False, visible=False)
                    with gr.Column(min_width=10):
                        with gr.Group(elem_id="share-btn-container") as share_group:
                            community_icon = gr.HTML(community_icon_html)
                            loading_icon = gr.HTML(loading_icon_html)
                            share_button = gr.Button("Share to community", elem_id="share-btn")
                gr.Markdown('<p style="font-size: 90%;opacity: 0.8;">Rating helps finding the most compatible LoRAs 🤗, results are shared annonymously <a href="https://huggingface.co/datasets/multimodalart/lora-fusing-preferences" target="_blank" rel="noopener noreferrer">here</a></p>')
  with gr.Accordion("Advanced settings", open=False):
    with gr.Row():
      lora_1_scale = gr.Slider(label="LoRA 1 Scale", minimum=0, maximum=1, step=0.1, value=0.7)
      lora_2_scale = gr.Slider(label="LoRa 2 Scale", minimum=0, maximum=1, step=0.1, value=0.7)
    negative_prompt = gr.Textbox(label="Negative prompt")
    seed = gr.Slider(label="Seed", info="-1 denotes a random seed", minimum=-1, maximum=2147483647, value=-1)
    last_used_seed = gr.Number(label="Last used seed", info="The seed used in the last generation", minimum=0, maximum=2147483647, value=-1, interactive=False)
  gr.Markdown("Generate with intent in [LoRA the Explorer](https://huggingface.co/spaces/multimodalart/LoraTheExplorer), but remember: sometimes restrictions flourish creativity 🌸")
    
  demo.load(shuffle_images, inputs=[], outputs=[lora_1_link, lora_1, lora_1_prompt, lora_1_id, lora_2_link, lora_2, lora_2_prompt, lora_2_id, prompt, shuffled_items, lora_1_scale, lora_2_scale], queue=False, show_progress="hidden")
  shuffle_button.click(shuffle_images, outputs=[lora_1_link, lora_1, lora_1_prompt, lora_1_id, lora_2_link, lora_2, lora_2_prompt, lora_2_id, prompt, shuffled_items, lora_1_scale, lora_2_scale], queue=False, show_progress="hidden")

  run_btn.click(hide_post_gen_info, outputs=[post_gen_info], queue=False).then(merge_and_run,
                inputs=[prompt, negative_prompt, shuffled_items, lora_1_scale, lora_2_scale, seed],
                outputs=[output_image, post_gen_info, last_used_seed, thumbs_up, thumbs_up_clicked, thumbs_down, thumbs_down_clicked],
                concurrency_limit=2)
  prompt.submit(hide_post_gen_info, outputs=[post_gen_info], queue=False).then(merge_and_run,
                inputs=[prompt, negative_prompt, shuffled_items, lora_1_scale, lora_2_scale, seed],
                outputs=[output_image, post_gen_info, last_used_seed, thumbs_up, thumbs_up_clicked, thumbs_down, thumbs_down_clicked],
                concurrency_limit=2)
  
  thumbs_up.click(save_preferences, inputs=[lora_1_id, lora_1_scale, lora_2_id, lora_2_scale, prompt, output_image, gr.State("up"), last_used_seed], outputs=[thumbs_up, thumbs_up_clicked, thumbs_down], queue=False)
  thumbs_down.click(save_preferences, inputs=[lora_1_id, lora_1_scale, lora_2_id, lora_2_scale, prompt, output_image, gr.State("down"), last_used_seed], outputs=[thumbs_down, thumbs_down_clicked, thumbs_up], queue=False)
  share_button.click(None, [], [], js=share_js)

demo.queue()
demo.launch()