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, FluxPipeline from live_preview_helpers import calculate_shift, retrieve_timesteps, flux_pipe_call_that_returns_an_iterable_of_images from diffusers.utils import load_image, export_to_gif from huggingface_hub import hf_hub_download, HfFileSystem, ModelCard, snapshot_download import copy import random import time import numpy as np from transformers import pipeline # Load translation model translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en") # 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" base_model = "black-forest-labs/FLUX.1-dev" 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) pipe = DiffusionPipeline.from_pretrained(base_model, torch_dtype=dtype, vae=taef1).to(device) 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 ) # Initialize FluxPipeline for GIF generation flux_pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to(device) MAX_SEED = 2**32-1 pipe.flux_pipe_call_that_returns_an_iterable_of_images = flux_pipe_call_that_returns_an_iterable_of_images.__get__(pipe) # Constants for GIF generation HEIGHT = 256 WIDTH = 1024 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 update_selection(evt: gr.SelectData, width, height): selected_lora = loras[evt.index] new_placeholder = f"Type a prompt for {selected_lora['title']}" lora_repo = selected_lora["repo"] updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✨" if "aspect" in selected_lora: if selected_lora["aspect"] == "portrait": width = 768 height = 1024 elif selected_lora["aspect"] == "landscape": width = 1024 height = 768 else: width = 1024 height = 1024 return ( gr.update(placeholder=new_placeholder), updated_text, evt.index, width, height, ) @spaces.GPU(duration=70) def generate_image(prompt_mash, steps, seed, cfg_scale, width, height, lora_scale, progress): 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": lora_scale}, output_type="pil", good_vae=good_vae, ): yield img @spaces.GPU(duration=70) def generate_image_to_image(prompt_mash, image_input_path, image_strength, steps, cfg_scale, width, height, lora_scale, seed): generator = torch.Generator(device="cuda").manual_seed(seed) pipe_i2i.to("cuda") 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": lora_scale}, output_type="pil", ).images[0] return final_image def translate_if_korean(text): # Check if the text contains Korean characters if any(ord(char) >= 0xAC00 and ord(char) <= 0xD7A3 for char in text): translated = translator(text, max_length=512) return translated[0]['translation_text'] return text def run_lora(prompt, image_input, image_strength, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale, progress=gr.Progress(track_tqdm=True)): if selected_index is None: raise gr.Error("You must select a LoRA before proceeding.") # Translate prompt if it contains Korean prompt = translate_if_korean(prompt) selected_lora = loras[selected_index] lora_path = selected_lora["repo"] trigger_word = selected_lora["trigger_word"] if(trigger_word): if "trigger_position" in selected_lora: if selected_lora["trigger_position"] == "prepend": prompt_mash = f"{trigger_word} {prompt}" else: prompt_mash = f"{prompt} {trigger_word}" else: prompt_mash = f"{trigger_word} {prompt}" else: prompt_mash = prompt with calculateDuration("Unloading LoRA"): pipe.unload_lora_weights() pipe_i2i.unload_lora_weights() # Load LoRA weights with calculateDuration(f"Loading LoRA weights for {selected_lora['title']}"): if(image_input is not None): if "weights" in selected_lora: pipe_i2i.load_lora_weights(lora_path, weight_name=selected_lora["weights"]) else: pipe_i2i.load_lora_weights(lora_path) else: if "weights" in selected_lora: pipe.load_lora_weights(lora_path, weight_name=selected_lora["weights"]) else: pipe.load_lora_weights(lora_path) # Set random seed for reproducibility 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, lora_scale, seed) yield final_image, seed, gr.update(visible=False) else: image_generator = generate_image(prompt_mash, steps, seed, cfg_scale, width, height, lora_scale, progress) # Consume the generator to get the final image 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) yield final_image, seed, gr.update(value=progress_bar, visible=False) 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(base_model) if((base_model != "black-forest-labs/FLUX.1-dev") and (base_model != "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() 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) gr.Warning(f"You didn't include a link neither a valid Hugging Face repository with a *.safetensors LoRA") raise Exception(f"You didn't include a link neither a valid Hugging Face repository with a *.safetensors LoRA") return split_link[1], link, safetensors_name, trigger_word, image_url def check_custom_model(link): if(link.startswith("https://")): if(link.startswith("https://huggingface.co") or link.startswith("https://www.huggingface.co")): link_split = link.split("huggingface.co/") return get_huggingface_safetensors(link_split[1]) else: return get_huggingface_safetensors(link) def add_custom_lora(custom_lora): global loras if(custom_lora): try: title, repo, path, trigger_word, image = check_custom_model(custom_lora) print(f"Loaded custom LoRA: {repo}") card = f'''
Loaded custom LoRA:

{title}

{"Using: "+trigger_word+" as the trigger word" if trigger_word else "No trigger word found. If there's a trigger word, include it in your prompt"}
''' existing_item_index = next((index for (index, item) in enumerate(loras) if item['repo'] == repo), None) if(not existing_item_index): new_item = { "image": image, "title": title, "repo": repo, "weights": path, "trigger_word": trigger_word } print(new_item) existing_item_index = len(loras) loras.append(new_item) return gr.update(visible=True, value=card), gr.update(visible=True), gr.Gallery(selected_index=None), f"Custom: {path}", existing_item_index, trigger_word except Exception as e: gr.Warning(f"Invalid LoRA: either you entered an invalid link, or a non-FLUX LoRA") return gr.update(visible=True, value=f"Invalid LoRA: either you entered an invalid link, a non-FLUX LoRA"), gr.update(visible=True), gr.update(), "", None, "" else: return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, "" def remove_custom_lora(): return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, "" run_lora.zerogpu = True # Functions for GIF generation def split_image(input_image, num_splits=4): output_images = [] for i in range(num_splits): left = i * 256 right = (i + 1) * 256 box = (left, 0, right, 256) output_images.append(input_image.crop(box)) return output_images def translate_to_english(text): return translator(text)[0]['translation_text'] @spaces.GPU() def predict(prompt, seed=42, randomize_seed=False, guidance_scale=5.0, num_inference_steps=28, progress=gr.Progress(track_tqdm=True)): if any('\u3131' <= char <= '\u318E' or '\uAC00' <= char <= '\uD7A3' for char in prompt): prompt = translate_to_english(prompt) prompt_template = f""" A side by side 4 frame image showing consecutive stills from a looped gif moving from left to right. The gif is of {prompt}. """ if randomize_seed: seed = random.randint(0, MAX_SEED) image = flux_pipe( prompt=prompt_template, guidance_scale=guidance_scale, num_inference_steps=num_inference_steps, num_images_per_prompt=1, generator=torch.Generator("cpu").manual_seed(seed), height=HEIGHT, width=WIDTH ).images[0] return export_to_gif(split_image(image, 4), "flux.gif", fps=4), image, seed 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: 0.5em} #gallery .grid-wrap{height: 10vh} #lora_list{background: var(--block-background-fill);padding: 0 1em .3em; font-size: 90%} .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} footer { visibility: hidden;} ''' with gr.Blocks(theme="Nymbo/Nymbo_Theme", css=css, delete_cache=(60, 3600)) as app: gr.HTML( """GINI 이미지 스튜디오 - 사용방법: 1) 원하는 '스타일' 클릭. 2) 프롬프트에 한글/영문 입력. 옵션) 원하는 스타일의 '이미지 업로드' 적용""", elem_id="title", ) with gr.Tabs(): with gr.TabItem("학습 모델"): selected_index = gr.State(None) with gr.Row(): with gr.Column(scale=3): prompt = gr.Textbox(label="Prompt (한글 또는 영어)", lines=1, placeholder="프롬프트를 입력하세요 (한글 입력 시 자동으로 영어로 번역됩니다)") with gr.Column(scale=1, elem_id="gen_column"): generate_button = gr.Button("생성", variant="primary", elem_id="gen_btn") with gr.Row(): with gr.Column(): selected_info = gr.Markdown("") gallery = gr.Gallery( [(item["image"], item["title"]) for item in loras], label="LoRA Gallery", allow_preview=False, columns=3, elem_id="gallery" ) with gr.Group(): custom_lora = gr.Textbox(label="Custom LoRA", info="LoRA Hugging Face path", placeholder="multimodalart/vintage-ads-flux") 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") custom_lora_info = gr.HTML(visible=False) custom_lora_button = gr.Button("Remove custom LoRA", visible=False) with gr.Column(): progress_bar = gr.Markdown(elem_id="progress",visible=False) result = gr.Image(label="Generated Image") with gr.Row(): with gr.Accordion("옵션", 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) lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=3, step=0.01, value=0.95) gallery.select( update_selection, inputs=[width, height], outputs=[prompt, selected_info, selected_index, width, height] ) custom_lora.input( add_custom_lora, inputs=[custom_lora], outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, prompt] ) custom_lora_button.click( remove_custom_lora, outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, custom_lora] ) gr.on( triggers=[generate_button.click, prompt.submit], fn=run_lora, inputs=[prompt, input_image, image_strength, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale], outputs=[result, seed, progress_bar] ) with gr.TabItem("GIF"): with gr.Column(elem_id="col-container"): with gr.Row(): prompt = gr.Text(label="프롬프트", show_label=False, max_lines=1, placeholder="프롬프트를 입력하세요") submit = gr.Button("제출", scale=0) output = gr.Image(label="GIF", show_label=False) output_stills = gr.Image(label="스틸 이미지", show_label=False, elem_id="stills") with gr.Accordion("고급 설정", open=False): seed = gr.Slider( label="시드", minimum=0, maximum=MAX_SEED, step=1, value=0, ) randomize_seed = gr.Checkbox(label="시드 무작위화", value=True) with gr.Row(): guidance_scale = gr.Slider( label="가이던스 스케일", minimum=1, maximum=15, step=0.1, value=3.5, ) num_inference_steps = gr.Slider( label="추론 단계 수", minimum=1, maximum=50, step=1, value=28, ) gr.Examples( examples=[ "고양이가 공중에서 발을 흔드는 모습", "팬더가 엉덩이를 좌우로 흔드는 모습", "꽃이 피어나는 과정" ], fn=predict, inputs=[prompt], outputs=[output, output_stills, seed], cache_examples="lazy" ) gr.on( triggers=[submit.click, prompt.submit], fn=predict, inputs=[prompt, seed, randomize_seed, guidance_scale, num_inference_steps], outputs=[output, output_stills, seed] ) app.queue() app.launch(auth=("gini","pick"))