import os import gradio as gr import json import logging import torch from PIL import Image import spaces from diffusers import DiffusionPipeline, AutoencoderTiny, AutoencoderKL, AutoPipelineForImage2Image from live_preview_helpers import calculate_shift, retrieve_timesteps, flux_pipe_call_that_returns_an_iterable_of_images from diffusers.utils import load_image from huggingface_hub import hf_hub_download, HfFileSystem, ModelCard, snapshot_download import copy import random import time import requests import pandas as pd from transformers import pipeline from gradio_imageslider import ImageSlider import numpy as np import warnings huggingface_token = os.getenv("HF_TOKEN") translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en", device="cpu") #Load prompts for randomization df = pd.read_csv('prompts.csv', header=None) prompt_values = df.values.flatten() # Load LoRAs from JSON file with open('loras.json', 'r') as f: loras = json.load(f) # Initialize the base model dtype = torch.bfloat16 device = "cuda" if torch.cuda.is_available() else "cpu" # 공통 FLUX 모델 로드 base_model = "black-forest-labs/FLUX.1-dev" pipe = DiffusionPipeline.from_pretrained(base_model, torch_dtype=dtype).to(device) # LoRA를 위한 설정 taef1 = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype).to(device) good_vae = AutoencoderKL.from_pretrained(base_model, subfolder="vae", torch_dtype=dtype).to(device) # Image-to-Image 파이프라인 설정 pipe_i2i = AutoPipelineForImage2Image.from_pretrained( base_model, vae=good_vae, transformer=pipe.transformer, text_encoder=pipe.text_encoder, tokenizer=pipe.tokenizer, text_encoder_2=pipe.text_encoder_2, tokenizer_2=pipe.tokenizer_2, torch_dtype=dtype ).to(device) MAX_SEED = 2**32 - 1 MAX_PIXEL_BUDGET = 1024 * 1024 pipe.flux_pipe_call_that_returns_an_iterable_of_images = flux_pipe_call_that_returns_an_iterable_of_images.__get__(pipe) class calculateDuration: def __init__(self, activity_name=""): self.activity_name = activity_name def __enter__(self): self.start_time = time.time() return self def __exit__(self, exc_type, exc_value, traceback): self.end_time = time.time() self.elapsed_time = self.end_time - self.start_time if self.activity_name: print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds") else: print(f"Elapsed time: {self.elapsed_time:.6f} seconds") def download_file(url, directory=None): if directory is None: directory = os.getcwd() # Use current working directory if not specified # Get the filename from the URL filename = url.split('/')[-1] # Full path for the downloaded file filepath = os.path.join(directory, filename) # Download the file response = requests.get(url) response.raise_for_status() # Raise an exception for bad status codes # Write the content to the file with open(filepath, 'wb') as file: file.write(response.content) return filepath def update_selection(evt: gr.SelectData, selected_indices, loras_state, width, height): selected_index = evt.index selected_indices = selected_indices or [] if selected_index in selected_indices: selected_indices.remove(selected_index) else: if len(selected_indices) < 3: selected_indices.append(selected_index) else: gr.Warning("You can select up to 3 LoRAs, remove one to select a new one.") return gr.update(), gr.update(), gr.update(), gr.update(), selected_indices, gr.update(), gr.update(), gr.update(), width, height, gr.update(), gr.update(), gr.update() selected_info_1 = "Select LoRA 1" selected_info_2 = "Select LoRA 2" selected_info_3 = "Select LoRA 3" lora_scale_1 = 1.15 lora_scale_2 = 1.15 lora_scale_3 = 1.15 lora_image_1 = None lora_image_2 = None lora_image_3 = None if len(selected_indices) >= 1: lora1 = loras_state[selected_indices[0]] selected_info_1 = f"### LoRA 1 Selected: [{lora1['title']}](https://huggingface.co/{lora1['repo']}) ✨" lora_image_1 = lora1['image'] if len(selected_indices) >= 2: lora2 = loras_state[selected_indices[1]] selected_info_2 = f"### LoRA 2 Selected: [{lora2['title']}](https://huggingface.co/{lora2['repo']}) ✨" lora_image_2 = lora2['image'] if len(selected_indices) >= 3: lora3 = loras_state[selected_indices[2]] selected_info_3 = f"### LoRA 3 Selected: [{lora3['title']}](https://huggingface.co/{lora3['repo']}) ✨" lora_image_3 = lora3['image'] if selected_indices: last_selected_lora = loras_state[selected_indices[-1]] new_placeholder = f"Type a prompt for {last_selected_lora['title']}" else: new_placeholder = "Type a prompt after selecting a LoRA" return gr.update(placeholder=new_placeholder), selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, width, height, lora_image_1, lora_image_2, lora_image_3 def remove_lora(selected_indices, loras_state, index_to_remove): if len(selected_indices) > index_to_remove: selected_indices.pop(index_to_remove) selected_info_1 = "Select LoRA 1" selected_info_2 = "Select LoRA 2" selected_info_3 = "Select LoRA 3" lora_scale_1 = 1.15 lora_scale_2 = 1.15 lora_scale_3 = 1.15 lora_image_1 = None lora_image_2 = None lora_image_3 = None for i, idx in enumerate(selected_indices): lora = loras_state[idx] if i == 0: selected_info_1 = f"### LoRA 1 Selected: [{lora['title']}]({lora['repo']}) ✨" lora_image_1 = lora['image'] elif i == 1: selected_info_2 = f"### LoRA 2 Selected: [{lora['title']}]({lora['repo']}) ✨" lora_image_2 = lora['image'] elif i == 2: selected_info_3 = f"### LoRA 3 Selected: [{lora['title']}]({lora['repo']}) ✨" lora_image_3 = lora['image'] return selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, lora_image_1, lora_image_2, lora_image_3 def remove_lora_1(selected_indices, loras_state): return remove_lora(selected_indices, loras_state, 0) def remove_lora_2(selected_indices, loras_state): return remove_lora(selected_indices, loras_state, 1) def remove_lora_3(selected_indices, loras_state): return remove_lora(selected_indices, loras_state, 2) def randomize_loras(selected_indices, loras_state): try: if len(loras_state) < 3: raise gr.Error("Not enough LoRAs to randomize.") selected_indices = random.sample(range(len(loras_state)), 3) lora1 = loras_state[selected_indices[0]] lora2 = loras_state[selected_indices[1]] lora3 = loras_state[selected_indices[2]] selected_info_1 = f"### LoRA 1 Selected: [{lora1['title']}](https://huggingface.co/{lora1['repo']}) ✨" selected_info_2 = f"### LoRA 2 Selected: [{lora2['title']}](https://huggingface.co/{lora2['repo']}) ✨" selected_info_3 = f"### LoRA 3 Selected: [{lora3['title']}](https://huggingface.co/{lora3['repo']}) ✨" lora_scale_1 = 1.15 lora_scale_2 = 1.15 lora_scale_3 = 1.15 lora_image_1 = lora1.get('image', 'path/to/default/image.png') lora_image_2 = lora2.get('image', 'path/to/default/image.png') lora_image_3 = lora3.get('image', 'path/to/default/image.png') random_prompt = random.choice(prompt_values) return selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, lora_image_1, lora_image_2, lora_image_3, random_prompt except Exception as e: print(f"Error in randomize_loras: {str(e)}") return "Error", "Error", "Error", [], 1.15, 1.15, 1.15, 'path/to/default/image.png', 'path/to/default/image.png', 'path/to/default/image.png', "" def add_custom_lora(custom_lora, selected_indices, current_loras): if custom_lora: try: title, repo, path, trigger_word, image = check_custom_model(custom_lora) print(f"Loaded custom LoRA: {repo}") existing_item_index = next((index for (index, item) in enumerate(current_loras) if item['repo'] == repo), None) if existing_item_index is None: if repo.endswith(".safetensors") and repo.startswith("http"): repo = download_file(repo) new_item = { "image": image if image else "/home/user/app/custom.png", "title": title, "repo": repo, "weights": path, "trigger_word": trigger_word } print(f"New LoRA: {new_item}") existing_item_index = len(current_loras) current_loras.append(new_item) # Update gallery gallery_items = [(item["image"], item["title"]) for item in current_loras] # Update selected_indices if there's room if len(selected_indices) < 3: selected_indices.append(existing_item_index) else: gr.Warning("You can select up to 3 LoRAs, remove one to select a new one.") # Update selected_info and images selected_info_1 = "Select a LoRA 1" selected_info_2 = "Select a LoRA 2" selected_info_3 = "Select a LoRA 3" lora_scale_1 = 1.15 lora_scale_2 = 1.15 lora_scale_3 = 1.15 lora_image_1 = None lora_image_2 = None lora_image_3 = None if len(selected_indices) >= 1: lora1 = current_loras[selected_indices[0]] selected_info_1 = f"### LoRA 1 Selected: {lora1['title']} ✨" lora_image_1 = lora1['image'] if lora1['image'] else None if len(selected_indices) >= 2: lora2 = current_loras[selected_indices[1]] selected_info_2 = f"### LoRA 2 Selected: {lora2['title']} ✨" lora_image_2 = lora2['image'] if lora2['image'] else None if len(selected_indices) >= 3: lora3 = current_loras[selected_indices[2]] selected_info_3 = f"### LoRA 3 Selected: {lora3['title']} ✨" lora_image_3 = lora3['image'] if lora3['image'] else None print("Finished adding custom LoRA") return ( current_loras, gr.update(value=gallery_items), selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, lora_image_1, lora_image_2, lora_image_3 ) except Exception as e: print(e) gr.Warning(str(e)) return current_loras, gr.update(), gr.update(), gr.update(), gr.update(), selected_indices, gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update() else: return current_loras, gr.update(), gr.update(), gr.update(), gr.update(), selected_indices, gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update() def remove_custom_lora(selected_indices, current_loras): if current_loras: custom_lora_repo = current_loras[-1]['repo'] # Remove from loras list current_loras = current_loras[:-1] # Remove from selected_indices if selected custom_lora_index = len(current_loras) if custom_lora_index in selected_indices: selected_indices.remove(custom_lora_index) # Update gallery gallery_items = [(item["image"], item["title"]) for item in current_loras] # Update selected_info and images selected_info_1 = "Select a LoRA 1" selected_info_2 = "Select a LoRA 2" selected_info_3 = "Select a LoRA 3" lora_scale_1 = 1.15 lora_scale_2 = 1.15 lora_scale_3 = 1.15 lora_image_1 = None lora_image_2 = None lora_image_3 = None if len(selected_indices) >= 1: lora1 = current_loras[selected_indices[0]] selected_info_1 = f"### LoRA 1 Selected: [{lora1['title']}]({lora1['repo']}) ✨" lora_image_1 = lora1['image'] if len(selected_indices) >= 2: lora2 = current_loras[selected_indices[1]] selected_info_2 = f"### LoRA 2 Selected: [{lora2['title']}]({lora2['repo']}) ✨" lora_image_2 = lora2['image'] if len(selected_indices) >= 3: lora3 = current_loras[selected_indices[2]] selected_info_3 = f"### LoRA 3 Selected: [{lora3['title']}]({lora3['repo']}) ✨" lora_image_3 = lora3['image'] return ( current_loras, gr.update(value=gallery_items), selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, lora_image_1, lora_image_2, lora_image_3 ) @spaces.GPU(duration=75) def generate_image(prompt_mash, steps, seed, cfg_scale, width, height, progress): print("Generating image...") pipe.to("cuda") generator = torch.Generator(device="cuda").manual_seed(seed) with calculateDuration("Generating image"): # Generate image for img in pipe.flux_pipe_call_that_returns_an_iterable_of_images( prompt=prompt_mash, num_inference_steps=steps, guidance_scale=cfg_scale, width=width, height=height, generator=generator, joint_attention_kwargs={"scale": 1.0}, output_type="pil", good_vae=good_vae, ): yield img @spaces.GPU(duration=75) def generate_image_to_image(prompt_mash, image_input_path, image_strength, steps, cfg_scale, width, height, seed): pipe_i2i.to("cuda") generator = torch.Generator(device="cuda").manual_seed(seed) image_input = load_image(image_input_path) final_image = pipe_i2i( prompt=prompt_mash, image=image_input, strength=image_strength, num_inference_steps=steps, guidance_scale=cfg_scale, width=width, height=height, generator=generator, joint_attention_kwargs={"scale": 1.0}, output_type="pil", ).images[0] return final_image def run_lora(prompt, image_input, image_strength, cfg_scale, steps, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, randomize_seed, seed, width, height, loras_state, progress=gr.Progress(track_tqdm=True)): try: # 한글 감지 및 번역 (이 부분은 그대로 유지) if any('\u3131' <= char <= '\u318E' or '\uAC00' <= char <= '\uD7A3' for char in prompt): translated = translator(prompt, max_length=512)[0]['translation_text'] print(f"Original prompt: {prompt}") print(f"Translated prompt: {translated}") prompt = translated if not selected_indices: raise gr.Error("You must select at least one LoRA before proceeding.") selected_loras = [loras_state[idx] for idx in selected_indices] # Build the prompt with trigger words (이 부분은 그대로 유지) prepends = [] appends = [] for lora in selected_loras: trigger_word = lora.get('trigger_word', '') if trigger_word: if lora.get("trigger_position") == "prepend": prepends.append(trigger_word) else: appends.append(trigger_word) prompt_mash = " ".join(prepends + [prompt] + appends) print("Prompt Mash: ", prompt_mash) # Unload previous LoRA weights with calculateDuration("Unloading LoRA"): pipe.unload_lora_weights() pipe_i2i.unload_lora_weights() print(f"Active adapters before loading: {pipe.get_active_adapters()}") # Load LoRA weights with respective scales lora_names = [] lora_weights = [] with calculateDuration("Loading LoRA weights"): for idx, lora in enumerate(selected_loras): try: lora_name = f"lora_{idx}" lora_path = lora['repo'] weight_name = lora.get("weights") print(f"Loading LoRA {lora_name} from {lora_path}") if image_input is not None: if weight_name: pipe_i2i.load_lora_weights(lora_path, weight_name=weight_name, adapter_name=lora_name) else: pipe_i2i.load_lora_weights(lora_path, adapter_name=lora_name) else: if weight_name: pipe.load_lora_weights(lora_path, weight_name=weight_name, adapter_name=lora_name) else: pipe.load_lora_weights(lora_path, adapter_name=lora_name) lora_names.append(lora_name) lora_weights.append(lora_scale_1 if idx == 0 else lora_scale_2 if idx == 1 else lora_scale_3) except Exception as e: print(f"Failed to load LoRA {lora_name}: {str(e)}") print("Loaded LoRAs:", lora_names) print("Adapter weights:", lora_weights) if lora_names: if image_input is not None: pipe_i2i.set_adapters(lora_names, adapter_weights=lora_weights) else: pipe.set_adapters(lora_names, adapter_weights=lora_weights) else: print("No LoRAs were successfully loaded.") return None, seed, gr.update(visible=False) print(f"Active adapters after loading: {pipe.get_active_adapters()}") # 여기서부터 이미지 생성 로직 (이 부분은 그대로 유지) with calculateDuration("Randomizing seed"): if randomize_seed: seed = random.randint(0, MAX_SEED) if image_input is not None: final_image = generate_image_to_image(prompt_mash, image_input, image_strength, steps, cfg_scale, width, height, seed) else: image_generator = generate_image(prompt_mash, steps, seed, cfg_scale, width, height, progress) final_image = None step_counter = 0 for image in image_generator: step_counter += 1 final_image = image progress_bar = f'
' yield image, seed, gr.update(value=progress_bar, visible=True) if final_image is None: raise Exception("Failed to generate image") return final_image, seed, gr.update(visible=False) except Exception as e: print(f"Error in run_lora: {str(e)}") return None, seed, gr.update(visible=False) run_lora.zerogpu = True def get_huggingface_safetensors(link): split_link = link.split("/") if len(split_link) == 2: model_card = ModelCard.load(link) base_model = model_card.data.get("base_model") print(f"Base model: {base_model}") if base_model not in ["black-forest-labs/FLUX.1-dev", "black-forest-labs/FLUX.1-schnell"]: raise Exception("Not a FLUX LoRA!") image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None) trigger_word = model_card.data.get("instance_prompt", "") image_url = f"https://huggingface.co/{link}/resolve/main/{image_path}" if image_path else None fs = HfFileSystem() safetensors_name = None try: list_of_files = fs.ls(link, detail=False) for file in list_of_files: if file.endswith(".safetensors"): safetensors_name = file.split("/")[-1] if not image_url and file.lower().endswith((".jpg", ".jpeg", ".png", ".webp")): image_elements = file.split("/") image_url = f"https://huggingface.co/{link}/resolve/main/{image_elements[-1]}" except Exception as e: print(e) raise gr.Error("Invalid Hugging Face repository with a *.safetensors LoRA") if not safetensors_name: raise gr.Error("No *.safetensors file found in the repository") return split_link[1], link, safetensors_name, trigger_word, image_url else: raise gr.Error("Invalid Hugging Face repository link") def check_custom_model(link): if link.endswith(".safetensors"): # Treat as direct link to the LoRA weights title = os.path.basename(link) repo = link path = None # No specific weight name trigger_word = "" image_url = None return title, repo, path, trigger_word, image_url elif link.startswith("https://"): if "huggingface.co" in link: link_split = link.split("huggingface.co/") return get_huggingface_safetensors(link_split[1]) else: raise Exception("Unsupported URL") else: # Assume it's a Hugging Face model path return get_huggingface_safetensors(link) def update_history(new_image, history): """Updates the history gallery with the new image.""" if history is None: history = [] if new_image is not None: history.insert(0, new_image) return history css = ''' #gen_btn{height: 100%} #title{text-align: center} #title h1{font-size: 3em; display:inline-flex; align-items:center} #title img{width: 100px; margin-right: 0.25em} #gallery .grid-wrap{height: 5vh} #lora_list{background: var(--block-background-fill);padding: 0 1em .3em; font-size: 90%} .custom_lora_card{margin-bottom: 1em} .card_internal{display: flex;height: 100px;margin-top: .5em} .card_internal img{margin-right: 1em} .styler{--form-gap-width: 0px !important} #progress{height:30px} #progress .generating{display:none} .progress-container {width: 100%;height: 30px;background-color: #f0f0f0;border-radius: 15px;overflow: hidden;margin-bottom: 20px} .progress-bar {height: 100%;background-color: #4f46e5;width: calc(var(--current) / var(--total) * 100%);transition: width 0.5s ease-in-out} #component-8, .button_total{height: 100%; align-self: stretch;} #loaded_loras [data-testid="block-info"]{font-size:80%} #custom_lora_structure{background: var(--block-background-fill)} #custom_lora_btn{margin-top: auto;margin-bottom: 11px} #random_btn{font-size: 300%} #component-11{align-self: stretch;} ''' #footer {visibility: hidden;} with gr.Blocks(theme="Nymbo/Nymbo_Theme", css=css, delete_cache=(60, 3600)) as app: loras_state = gr.State(loras) selected_indices = gr.State([]) gr.Markdown( """ # MixGen3: 멀티 Lora(이미지 학습) 통합 생성 모델 ### 사용 안내: 1) 갤러리에서 원하는 모델을 선택(최대 3개까지) 2) 프롬프트에 한글 또는 영문으로 원하는 내용을 입력 3) Generate 버튼 실행 ### Contacts: arxivgpt@gmail.com """ ) with gr.Tab(label="Generate"): with gr.Row(): with gr.Column(scale=3): prompt = gr.Textbox(label="Prompt", lines=1, placeholder="Type a prompt after selecting a LoRA") with gr.Column(scale=1): generate_button = gr.Button("Generate", variant="primary", elem_classes=["button_total"]) with gr.Row(elem_id="loaded_loras"): with gr.Column(scale=1, min_width=25): randomize_button = gr.Button("🎲", variant="secondary", scale=1, elem_id="random_btn") with gr.Column(scale=8): with gr.Row(): with gr.Column(scale=0, min_width=50): lora_image_1 = gr.Image(label="LoRA 1 Image", interactive=False, min_width=50, width=50, show_label=False, show_share_button=False, show_download_button=False, show_fullscreen_button=False, height=50) with gr.Column(scale=3, min_width=100): selected_info_1 = gr.Markdown("Select a LoRA 1") with gr.Column(scale=5, min_width=50): lora_scale_1 = gr.Slider(label="LoRA 1 Scale", minimum=0, maximum=3, step=0.01, value=1.15) with gr.Row(): remove_button_1 = gr.Button("Remove", size="sm") with gr.Column(scale=8): with gr.Row(): with gr.Column(scale=0, min_width=50): lora_image_2 = gr.Image(label="LoRA 2 Image", interactive=False, min_width=50, width=50, show_label=False, show_share_button=False, show_download_button=False, show_fullscreen_button=False, height=50) with gr.Column(scale=3, min_width=100): selected_info_2 = gr.Markdown("Select a LoRA 2") with gr.Column(scale=5, min_width=50): lora_scale_2 = gr.Slider(label="LoRA 2 Scale", minimum=0, maximum=3, step=0.01, value=1.15) with gr.Row(): remove_button_2 = gr.Button("Remove", size="sm") with gr.Column(scale=8): with gr.Row(): with gr.Column(scale=0, min_width=50): lora_image_3 = gr.Image(label="LoRA 3 Image", interactive=False, min_width=50, width=50, show_label=False, show_share_button=False, show_download_button=False, show_fullscreen_button=False, height=50) with gr.Column(scale=3, min_width=100): selected_info_3 = gr.Markdown("Select a LoRA 3") with gr.Column(scale=5, min_width=50): lora_scale_3 = gr.Slider(label="LoRA 3 Scale", minimum=0, maximum=3, step=0.01, value=1.15) with gr.Row(): remove_button_3 = gr.Button("Remove", size="sm") with gr.Row(): with gr.Column(): with gr.Group(): with gr.Row(elem_id="custom_lora_structure"): custom_lora = gr.Textbox(label="Custom LoRA", info="LoRA Hugging Face path or *.safetensors public URL", placeholder="ginipick/flux-lora-eric-cat", scale=3, min_width=150) add_custom_lora_button = gr.Button("Add Custom LoRA", elem_id="custom_lora_btn", scale=2, min_width=150) remove_custom_lora_button = gr.Button("Remove Custom LoRA", visible=False) gr.Markdown("[Check the list of FLUX LoRAs](https://huggingface.co/models?other=base_model:adapter:black-forest-labs/FLUX.1-dev)", elem_id="lora_list") gallery = gr.Gallery( [(item["image"], item["title"]) for item in loras], label="Or pick from the LoRA Explorer gallery", allow_preview=False, columns=4, elem_id="gallery" ) with gr.Column(): progress_bar = gr.Markdown(elem_id="progress", visible=False) result = gr.Image(label="Generated Image", interactive=False) with gr.Accordion("History", open=False): history_gallery = gr.Gallery(label="History", columns=6, object_fit="contain", interactive=False) with gr.Row(): with gr.Accordion("Advanced Settings", open=False): with gr.Row(): input_image = gr.Image(label="Input image", type="filepath") image_strength = gr.Slider(label="Denoise Strength", info="Lower means more image influence", minimum=0.1, maximum=1.0, step=0.01, value=0.75) with gr.Column(): with gr.Row(): cfg_scale = gr.Slider(label="CFG Scale", minimum=1, maximum=20, step=0.5, value=3.5) steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=28) with gr.Row(): width = gr.Slider(label="Width", minimum=256, maximum=1536, step=64, value=1024) height = gr.Slider(label="Height", minimum=256, maximum=1536, step=64, value=1024) with gr.Row(): randomize_seed = gr.Checkbox(True, label="Randomize seed") seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True) with gr.Tab(label="Guide"): gr.Markdown( """ ### Guide # MixGen3: Next-Generation Image Generation Service, Embracing the Charm of LoRA ## The New Paradigm of Image Generation With the advancement of artificial intelligence, image generation technology is evolving at an astonishing speed. In particular, text-to-image generation models, which create desired images from text input, are gaining significant attention in creative content production and the art field. Recently, thanks to the efforts of various open-source communities and researchers, image generation models that anyone can easily access have emerged. Among them, the LoRA (Low-Rank Adaptation) model is attracting attention as a technology that presents a new paradigm in image generation. ## The Encounter of LoRA and MixGen3 ### The Emergence of LoRA Models LoRA models are lightweight models that adapt pre-trained large AI models to generate new styles or content. LoRA has the advantage of efficient image generation with small capacity and the ability to express various styles. Through open-source platforms like Hugging Face, various LoRA models are being shared and researched, promoting the development of the image generation field. ### Introduction to MixGen3 Service MixGen3, which we'll introduce in this blog, is an innovative service that generates images using these LoRA models. Through MixGen3, users can select various LoRA models and input their own prompts to create original images. Now, let's take a closer look at the features and value of the MixGen3 service. ## How to Use MixGen3 Image Generation Service This service is a space where you can fully enjoy the charm of LoRA (Low-Rank Adaptation) models. LoRA models are AI models that generate images and can express various styles and content. Now, let's dive into the world of LoRA models through MixGen3! ### Service Screen Configuration The MixGen3 service screen is divided into four main areas: the gallery located on the left side of the screen, the prompt input box and Generate button in the center, the image size adjustment slider and advanced settings at the bottom, and the menu button in the upper right corner of the screen. - Gallery: The gallery on the left side of the service screen is where you select LoRA models. Various LoRA models are displayed with images. Scroll through the gallery to find the LoRA model you want. - Prompt Input Box: The prompt input box is where you enter the content of the image you want to generate. You can input in Korean or English what you want to express in the image. Use your creativity to imagine the image you want! - Generate Button: The Generate button starts image generation. After entering the prompt, click this button to start generating the image. Now the LoRA model will turn your imagination into reality! - Image Size Adjustment: The image size adjustment is a slider that adjusts the size of the image to be generated. You can set the image ratio by adjusting the width and height as desired. - Advanced Settings: By clicking the advanced settings button, you can adjust detailed options for image generation. You can create more sophisticated images by adjusting image intensity, CFG scale, number of steps, etc. - Menu Button: By clicking the menu button in the upper right corner of the screen, you can access various menus such as service settings, help, and sending feedback. ### Selecting LoRA Models Selecting LoRA models is the most important process in the service. That's because LoRA models are the key elements that determine the style and content of the image! Now, let's select a LoRA model. 1. Find the LoRA model you want in the gallery. The gallery displays images and titles of LoRA models. Each LoRA model has a unique style, so choose the one you like. 2. Click on a LoRA model with your mouse, and the selected LoRA model will be added to the selected LoRA list at the top of the screen. You can select up to 3 LoRA models. 3. In the selected LoRA list, you can change the order of LoRA models or adjust the scale. The scale adjusts the influence of the LoRA model, so move the slider to set the desired value. ### Entering Prompts and Generating Images Now that we've selected the LoRA model, shall we enter prompts and generate images? 1. Enter the content you want to express in the image in the prompt input box. For example, try expressing the image you imagine in words, such as "beautiful forest scenery", "cute cat family", "surreal city". 2. When you click the Generate button, image generation begins. A progress bar appears, showing the process of the LoRA model generating the image. 3. After a short wait, the generated image appears in the center of the screen. The LoRA model will have implemented your imagination into a wonderful image! ### Checking and Saving Image Results Here's how to check and save the generated image: 1. The generated image is displayed large in the center of the screen. By clicking on the image, you can enlarge it to full screen to see it in detail. 2. By clicking the download button at the bottom right of the image, you can download the image. Save the image where you want so you can view it again later. 3. You can change the size of the image using the image size adjustment slider. If you want to save the image in a different size, adjust it to the desired size and then download it. ### Random Seed and Advanced Settings MixGen3 provides various features for more creative image generation. Let's look at the random seed and advanced settings features. - Random Seed: Random seed is a feature for the diversity of image generation. By clicking the 'Random Seed' button, a new seed is generated each time, allowing you to create slightly different images. Explore various images! - Advanced Settings: By clicking the advanced settings button, you can adjust detailed options for image generation. You can create more sophisticated images by increasing image intensity or adjusting the number of steps. Try to maximize the capabilities of the LoRA model through advanced settings! ### Gallery and Help MixGen3 provides various features for user convenience. You can use the service more easily through the gallery and help. - Gallery: By clicking the gallery button at the bottom of the service screen, you can check the images you've generated previously. Find images you want to see again and compare the effects of LoRA models. ## Conclusion We've now learned in detail how to use the MixGen3 image generation service. Select LoRA models, enter prompts, and generate your own original images. If you explore various LoRA models and use your creativity, you can get amazing results. ### Differentiation from Existing Services: 1. Integration of LoRA Models: MixGen3 integrates multiple LoRA models into a single service. Users can explore various LoRA models through the gallery and select the desired models. This is a much more convenient and efficient approach compared to the existing method of having to search for and use individual LoRA models. 2. Combination of Various LoRA Models: The ability to combine up to 3 LoRA models to generate images is MixGen3's biggest differentiator. By fusing multiple LoRA models, you can express unique styles and content that didn't exist before. For example, you can combine a landscape LoRA with an animal LoRA to create an image of a dog playing on a beautiful beach. 3. User-Friendly Interface: MixGen3 provides an intuitive interface considering user convenience. You can easily select LoRA models from the gallery, enter prompts, and generate images with just one button click. Also, it supports user creativity through various options such as image size adjustment, random seed, and advanced settings. ### Applications of MixGen3: 1. Content Creation: MixGen3 can be an inspiring tool for writers, artists, and content creators. You can visualize various ideas through prompts and create original works by combining LoRA models. 2. Design and Illustration: Graphic designers or illustrators can quickly visualize ideas through MixGen3, explore various styles of LoRA models, and complete unique designs. 3. Marketing and Advertising: Marketers can use MixGen3 to generate novel advertising images and strengthen brand images through various LoRA models. 4. Education and Learning: In the field of education, MixGen3 can be a tool to enhance students' creativity and help with visual learning. It can express various topics such as history, science, and art in images, adding interest to learning. ### The Value of MixGen3: 1. Enhancing Creativity: MixGen3 maximizes users' creativity. Through the combination of various LoRA models and prompts, users can unleash their imagination and generate original images. 2. Time Saving: You can save time spent finding and applying LoRA models directly. MixGen3 provides verified LoRA models and makes image generation simple with a convenient interface. 3. Possibility of Collaboration: MixGen3 provides a platform for multiple users to generate images together and share opinions. Creators working on joint projects can share LoRA models, give and receive feedback, and collaborate. 4. Continuous Development: MixGen3 utilizes LoRA models from the open-source community and contributes to their development. It improves the service and LoRA models through user feedback and suggestions, promoting the development of the image generation field. ### Expected Effects: 1. Diversity of Content: Images generated through MixGen3 will increase the diversity of online content. By using images reflecting the unique style of LoRA models in blogs, social media, websites, etc., differentiated content can be produced. 2. Lowering the Entry Barrier for Creation: Even people without professional design or artistic skills can easily challenge image generation through MixGen3. This lowers the entry barrier for creative activities and helps more people enjoy creative activities. 3. Improving Creativity: MixGen3 stimulates users' creativity and helps explore new ideas. The combination of various LoRA models induces creative thinking and provides new inspiration. 4. Improving Productivity: Using MixGen3 simplifies the image generation process and saves time. This increases the productivity of content creation and supports faster testing of more ideas. ## Conclusion: Spreading the Wings of Imagination MixGen3 brings a new wind to the field of image generation by fully utilizing the charm of LoRA models. Users can combine various LoRA models, unleash their imagination through prompts, and generate original images. Content creators, designers, marketers, and experts in various fields can maximize their creativity and experience efficient image generation through MixGen3. Now, try MixGen3 yourself! Visit https://openfree-mixgen3.hf.space and explore various LoRA models and enter prompts for free. You will have an amazing experience of your imagined images becoming reality. We hope you spread the wings of creativity with MixGen3 and enjoy the pleasure of image generation! For any further questions, opinions, or suggestions, please contact us anytime at arxivgpt@gmail.com. """ ) with gr.Tab(label="FAQ"): gr.Markdown( """ ### Frequently Asked Questions (FAQ) 1. What is MixGen3? MixGen3 is an online service that generates images using LoRA (Low-Rank Adaptation) models. Users can select desired LoRA models, input prompts, and generate images that match the style and content of the chosen LoRA models. You can combine various LoRA models to create creative and unique images. 2. What are LoRA models? LoRA (Low-Rank Adaptation) models are lightweight models used in AI image generation. LoRA is added to pre-trained large models to adapt them to generate new styles or content. LoRA has the advantage of efficiently generating images with a small capacity and expressing various styles. 3. How do I select LoRA models from the gallery? You can select LoRA models from the gallery located on the left side of the service screen. The gallery displays various LoRA models with images, titles, and brief descriptions. You can click to select up to 3 LoRA models. The selected LoRA models are displayed in the selected LoRA list at the top of the screen. 4. How many LoRA models can I select? Currently, in the MixGen3 service, you can select up to 3 LoRA models. You can generate images that fuse various styles and content by combining multiple LoRA models. After selecting LoRA models, you can adjust the scale to control the influence of each model. 5. How do I input prompts? Prompts are entered in Korean or English in the input box located in the center of the service screen. A prompt is text that describes the content or style of the image to be generated. For example, you can enter simple words or sentences like "beautiful beach", "cute puppy", or "surreal landscape". The more specific and detailed the prompt, the closer the result will be to the desired image. 6. Should I include trigger words for LoRA models in the prompt? Each LoRA model has its own trigger words. Trigger words activate the corresponding LoRA model to induce the generation of specific styles or content. Including these trigger words in the prompt can result in images that better reflect the characteristics of the LoRA model. However, it's not mandatory to include trigger words. LoRA models can generate images with general words, but using trigger words is recommended for more specific results. 7. What happens when I click the Generate button? When you click the Generate button, an image is created based on the selected LoRA models and prompt. This process goes through several stages. First, a draft of the image is generated based on the prompt. Then, the selected LoRA models are sequentially applied to modify and enhance the image. This process is displayed with a progress bar, and when completed, the final image appears in the result window. 8. How long does it take to generate an image? The image generation time varies depending on the number of selected LoRA models, image size, and advanced settings. Generally, when selecting 3 LoRA models and generating with the default image size, it takes about 1 to 2 minutes. Adjusting advanced settings such as image intensity and step count can increase generation time. 9. Where are the generated images saved? Generated images are saved in the gallery within the service. You can view previous generated images in the gallery and download them again. Also, after generating an image, a button to download the image file is provided in the result window. You can click this button to download the image and save it where you want. 10. How do I adjust the image size? You can adjust the image size using the slider at the bottom of the service screen. You can set the desired image ratio by adjusting the width and height separately. Image size affects image generation, so generating higher resolution images may take more time. 11. What is a random seed and how is it used? A random seed is a function that controls the randomness of image generation. The seed determines the diversity of the image, so even with the same prompt and LoRA model, applying a different seed will generate slightly different images. Clicking the 'Random Seed' button generates a new seed each time, allowing you to explore various images. 12. What are the advanced settings? In advanced settings, you can adjust detailed options for image generation. This includes image strength, CFG scale, number of steps, etc. Image strength adjusts the influence of the LoRA model, CFG scale adjusts the detail expression of the image, and the number of steps refers to the number of iterations in the image generation process. Advanced settings allow for more sophisticated and detailed image generation. 13. Can I add LoRA models directly? Currently, the MixGen3 service does not provide a function to directly add LoRA models. The service operates based on LoRA models provided by Hugging Face, and you can select from LoRA models available on that platform. We plan to continuously update with various LoRA models that users prefer. If you want to train your own image, we can help. Please contact us anytime at arxivgpt@gmail.com. 14. How are the LoRA models provided in the service selected? The LoRA models included in the service are selected from popular and well-known models on the Hugging Face platform. We also plan to regularly update considering the quality and diversity of LoRA models. If users have LoRA models they want, please request them through feedback, and we will consider reflecting them. 15. Can I use the generated images commercially? Images generated by MixGen3 are suitable for personal or non-commercial use. If you want to use them commercially, you need to consider copyright and license issues. As each LoRA model may have different licenses, please check the license information for each model. For detailed inquiries about commercial use, please contact the service operator. 16. Is the service available on mobile devices? The MixGen3 service is optimized for use on mobile devices. You can access the service through a mobile web browser to generate images and explore the gallery. It provides a user experience similar to desktop on mobile devices. 17. What should I be cautious about when using the service? MixGen3 is a service developed to support users' creative activities. Therefore, care should be taken not to infringe on others' copyrights or portrait rights with generated images. Also, excessive image generation can burden the server, so please use it at appropriate intervals. 18. What should I do if image generation fails? If image generation fails, you can try a few solutions. First, try modifying the prompt to enter clearer and more specific content. Also, adjusting the scale of LoRA models or selecting different LoRA models can help. If it continues to fail, please contact the service operator for technical support. 19. What are the service update plans? The MixGen3 service will be improved and developed through continuous updates. We are planning to add LoRA models, enhance the user interface, introduce new features, etc. Update content will be shared through service notices or social media, so please stay tuned. 20. How can I provide feedback or suggestions? Feedback or suggestions for the MixGen3 service are always welcome! Please leave your opinion by clicking the feedback button in the service, or contact us via email or social media. Your valuable opinions are a great help in improving the service. We will always keep our ears open to provide better service. """ ) gallery.select( update_selection, inputs=[selected_indices, loras_state, width, height], outputs=[prompt, selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, width, height, lora_image_1, lora_image_2, lora_image_3] ) remove_button_1.click( remove_lora_1, inputs=[selected_indices, loras_state], outputs=[selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, lora_image_1, lora_image_2, lora_image_3] ) remove_button_2.click( remove_lora_2, inputs=[selected_indices, loras_state], outputs=[selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, lora_image_1, lora_image_2, lora_image_3] ) remove_button_3.click( remove_lora_3, inputs=[selected_indices, loras_state], outputs=[selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, lora_image_1, lora_image_2, lora_image_3] ) randomize_button.click( randomize_loras, inputs=[selected_indices, loras_state], outputs=[selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, lora_image_1, lora_image_2, lora_image_3, prompt] ) add_custom_lora_button.click( add_custom_lora, inputs=[custom_lora, selected_indices, loras_state], outputs=[loras_state, gallery, selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, lora_image_1, lora_image_2, lora_image_3] ) remove_custom_lora_button.click( remove_custom_lora, inputs=[selected_indices, loras_state], outputs=[loras_state, gallery, selected_info_1, selected_info_2, selected_info_3, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, lora_image_1, lora_image_2, lora_image_3] ) gr.on( triggers=[generate_button.click, prompt.submit], fn=run_lora, inputs=[prompt, input_image, image_strength, cfg_scale, steps, selected_indices, lora_scale_1, lora_scale_2, lora_scale_3, randomize_seed, seed, width, height, loras_state], outputs=[result, seed, progress_bar] ).then( fn=lambda x, history: update_history(x, history) if x is not None else history, inputs=[result, history_gallery], outputs=history_gallery, ) if __name__ == "__main__": app.queue(max_size=20) app.launch(debug=True)