import gradio as gr import torch from huggingface_hub import hf_hub_download from safetensors.torch import load_file from share_btn import community_icon_html, loading_icon_html, share_js from cog_sdxl_dataset_and_utils import TokenEmbeddingsHandler import lora import copy import json import gc import random from urllib.parse import quote import gdown import os import diffusers from diffusers.utils import load_image from diffusers.models import ControlNetModel from diffusers import AutoencoderKL, DPMSolverMultistepScheduler import cv2 import torch import numpy as np from PIL import Image from insightface.app import FaceAnalysis from pipeline_stable_diffusion_xl_instantid_img2img import StableDiffusionXLInstantIDImg2ImgPipeline, draw_kps from controlnet_aux import ZoeDetector with open("sdxl_loras.json", "r") as file: data = json.load(file) sdxl_loras_raw = [ { "image": 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), "likes": item.get("likes", 0), "downloads": item.get("downloads", 0), "is_nc": item.get("is_nc", False), "new": item.get("new", False), } for item in data ] device = "cuda" state_dicts = {} for item in sdxl_loras_raw: saved_name = hf_hub_download(item["repo"], item["weights"]) if not saved_name.endswith('.safetensors'): state_dict = torch.load(saved_name) else: state_dict = load_file(saved_name) state_dicts[item["repo"]] = { "saved_name": saved_name, "state_dict": state_dict } sdxl_loras_raw_new = [item for item in sdxl_loras_raw if item.get("new") == True] sdxl_loras_raw = [item for item in sdxl_loras_raw if item.get("new") != True] # download models hf_hub_download( repo_id="InstantX/InstantID", filename="ControlNetModel/config.json", local_dir="/data/checkpoints", ) hf_hub_download( repo_id="InstantX/InstantID", filename="ControlNetModel/diffusion_pytorch_model.safetensors", local_dir="/data/checkpoints", ) hf_hub_download( repo_id="InstantX/InstantID", filename="ip-adapter.bin", local_dir="/data/checkpoints" ) hf_hub_download( repo_id="latent-consistency/lcm-lora-sdxl", filename="pytorch_lora_weights.safetensors", local_dir="/data/checkpoints", ) # download antelopev2 gdown.download(url="https://drive.google.com/file/d/18wEUfMNohBJ4K3Ly5wpTejPfDzp-8fI8/view?usp=sharing", output="/data/", quiet=False, fuzzy=True) # unzip antelopev2.zip os.system("unzip /data/antelopev2.zip -d /data/models/") app = FaceAnalysis(name='antelopev2', root='/data', providers=['CPUExecutionProvider']) app.prepare(ctx_id=0, det_size=(640, 640)) # prepare models under ./checkpoints face_adapter = f'/data/checkpoints/ip-adapter.bin' controlnet_path = f'/data/checkpoints/ControlNetModel' # load IdentityNet identitynet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16) zoedepthnet = ControlNetModel.from_pretrained("diffusers/controlnet-zoe-depth-sdxl-1.0",torch_dtype=torch.float16) vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16) pipe = StableDiffusionXLInstantIDImg2ImgPipeline.from_pretrained("rubbrband/albedobaseXL_v21", vae=vae, controlnet=[identitynet, zoedepthnet], torch_dtype=torch.float16) pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True) pipe.load_ip_adapter_instantid(face_adapter) pipe.set_ip_adapter_scale(0.8) zoe = ZoeDetector.from_pretrained("lllyasviel/Annotators") zoe.to("cuda") original_pipe = copy.deepcopy(pipe) pipe.to(device) last_lora = "" last_merged = False last_fused = False js = ''' var button = document.getElementById('button'); // Add a click event listener to the button button.addEventListener('click', function() { element.classList.add('selected'); }); ''' def update_selection(selected_state: gr.SelectData, sdxl_loras, is_new=False): lora_repo = sdxl_loras[selected_state.index]["repo"] instance_prompt = sdxl_loras[selected_state.index]["trigger_word"] new_placeholder = "Type a prompt. This LoRA applies for all prompts, no need for a trigger word" if instance_prompt == "" else "Type a prompt to use your selected LoRA" weight_name = sdxl_loras[selected_state.index]["weights"] updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✨ {'(non-commercial LoRA, `cc-by-nc`)' if sdxl_loras[selected_state.index]['is_nc'] else '' }" is_compatible = sdxl_loras[selected_state.index]["is_compatible"] is_pivotal = sdxl_loras[selected_state.index]["is_pivotal"] use_with_diffusers = f''' ## Using [`{lora_repo}`](https://huggingface.co/{lora_repo}) ## Use it with diffusers: ''' if is_compatible: use_with_diffusers += f''' from diffusers import StableDiffusionXLPipeline import torch model_path = "stabilityai/stable-diffusion-xl-base-1.0" pipe = StableDiffusionPipeline.from_pretrained(model_path, torch_dtype=torch.float16) pipe.to("cuda") pipe.load_lora_weights("{lora_repo}", weight_name="{weight_name}") prompt = "{instance_prompt}..." lora_scale= 0.9 image = pipe(prompt, num_inference_steps=30, guidance_scale=7.5, cross_attention_kwargs={{"scale": lora_scale}}).images[0] image.save("image.png") ''' elif not is_pivotal: use_with_diffusers += "This LoRA is not compatible with diffusers natively yet. But you can still use it on diffusers with `bmaltais/kohya_ss` LoRA class, check out this [Google Colab](https://colab.research.google.com/drive/14aEJsKdEQ9_kyfsiV6JDok799kxPul0j )" else: use_with_diffusers += f"This LoRA is not compatible with diffusers natively yet. But you can still use it on diffusers with sdxl-cog `TokenEmbeddingsHandler` class, check out the [model repo](https://huggingface.co/{lora_repo}#inference-with-🧨-diffusers)" use_with_uis = f''' ## Use it with Comfy UI, Invoke AI, SD.Next, AUTO1111: ### Download the `*.safetensors` weights of [here](https://huggingface.co/{lora_repo}/resolve/main/{weight_name}) - [ComfyUI guide](https://comfyanonymous.github.io/ComfyUI_examples/lora/) - [Invoke AI guide](https://invoke-ai.github.io/InvokeAI/features/CONCEPTS/?h=lora#using-loras) - [SD.Next guide](https://github.com/vladmandic/automatic) - [AUTOMATIC1111 guide](https://stable-diffusion-art.com/lora/) ''' if(is_new): if(selected_state.index == 0): selected_state.index = -9999 else: selected_state.index *= -1 return ( updated_text, instance_prompt, gr.update(placeholder=new_placeholder), selected_state, use_with_diffusers, use_with_uis, gr.Gallery(selected_index=None) ) def check_selected(selected_state): if not selected_state: raise gr.Error("You must select a LoRA") def merge_incompatible_lora(full_path_lora, lora_scale): for weights_file in [full_path_lora]: if ";" in weights_file: weights_file, multiplier = weights_file.split(";") multiplier = float(multiplier) else: multiplier = lora_scale lora_model, weights_sd = lora.create_network_from_weights( multiplier, full_path_lora, pipe.vae, pipe.text_encoder, pipe.unet, for_inference=True, ) lora_model.merge_to( pipe.text_encoder, pipe.unet, weights_sd, torch.float16, "cuda" ) del weights_sd del lora_model gc.collect() def run_lora(face_image, prompt, negative, lora_scale, selected_state, face_strength, image_strength, guidance_scale, depth_control_scale, sdxl_loras, sdxl_loras_new, progress=gr.Progress(track_tqdm=True)): global last_lora, last_merged, last_fused, pipe face_info = app.get(cv2.cvtColor(np.array(face_image), cv2.COLOR_RGB2BGR)) face_info = sorted(face_info, key=lambda x:(x['bbox'][2]-x['bbox'][0])*x['bbox'][3]-x['bbox'][1])[-1] # only use the maximum face face_emb = face_info['embedding'] face_kps = draw_kps(face_image, face_info['kps']) #prepare face zoe with torch.no_grad(): image_zoe = zoe(face_image) width, height = face_kps.size images = [face_kps, image_zoe.resize((height, width))] if(selected_state.index < 0): if(selected_state.index == -9999): selected_state.index = 0 else: selected_state.index *= -1 sdxl_loras = sdxl_loras_new print("Selected State: ", selected_state.index) print(sdxl_loras[selected_state.index]["repo"]) if negative == "": negative = None if not selected_state: raise gr.Error("You must select a LoRA") repo_name = sdxl_loras[selected_state.index]["repo"] weight_name = sdxl_loras[selected_state.index]["weights"] full_path_lora = state_dicts[repo_name]["saved_name"] loaded_state_dict = copy.deepcopy(state_dicts[repo_name]["state_dict"]) cross_attention_kwargs = None if last_lora != repo_name: if(last_fused): pipe.unfuse_lora() pipe.load_lora_weights(loaded_state_dict) pipe.fuse_lora() last_fused = True is_pivotal = sdxl_loras[selected_state.index]["is_pivotal"] if(is_pivotal): #Add the textual inversion embeddings from pivotal tuning models text_embedding_name = sdxl_loras[selected_state.index]["text_embedding_weights"] embedding_path = hf_hub_download(repo_id=repo_name, filename=text_embedding_name, repo_type="model") state_dict_embedding = load_file(embedding_path) print(state_dict_embedding) try: pipe.unload_textual_inversion() pipe.load_textual_inversion(state_dict_embedding["clip_l"], token=["", ""], text_encoder=pipe.text_encoder, tokenizer=pipe.tokenizer) pipe.load_textual_inversion(state_dict_embedding["clip_g"], token=["", ""], text_encoder=pipe.text_encoder_2, tokenizer=pipe.tokenizer_2) except: pipe.unload_textual_inversion() pipe.load_textual_inversion(state_dict_embedding["text_encoders_0"], token=["", ""], text_encoder=pipe.text_encoder, tokenizer=pipe.tokenizer) pipe.load_textual_inversion(state_dict_embedding["text_encoders_1"], token=["", ""], text_encoder=pipe.text_encoder_2, tokenizer=pipe.tokenizer_2) image = pipe( prompt=prompt, negative_prompt=negative, width=1024, height=1024, image_embeds=face_emb, image=face_image, strength=1-image_strength, control_image=images, num_inference_steps=20, guidance_scale = 7, controlnet_conditioning_scale=[face_strength, 0.8], ).images[0] last_lora = repo_name gc.collect() return image, gr.update(visible=True) def shuffle_gallery(sdxl_loras): random.shuffle(sdxl_loras) return [(item["image"], item["title"]) for item in sdxl_loras], sdxl_loras def swap_gallery(order, sdxl_loras): if(order == "random"): return shuffle_gallery(sdxl_loras) else: sorted_gallery = sorted(sdxl_loras, key=lambda x: x.get(order, 0), reverse=True) return [(item["image"], item["title"]) for item in sorted_gallery], sorted_gallery def deselect(): return gr.Gallery(selected_index=None) with gr.Blocks(css="custom.css") as demo: gr_sdxl_loras = gr.State(value=sdxl_loras_raw) gr_sdxl_loras_new = gr.State(value=sdxl_loras_raw_new) title = gr.HTML( """

LoRA Face to All

""", elem_id="title", ) selected_state = gr.State() with gr.Row(elem_id="main_app"): with gr.Group(elem_id="gallery_box"): photo = gr.Image(label="Upload a picture of yourself", interactive=True, type="pil") selected_loras = gr.Gallery(label="Selected LoRAs", height=80, show_share_button=False, visible=False, elem_id="gallery_selected", ) order_gallery = gr.Radio(choices=["random", "likes"], value="random", label="Order by", elem_id="order_radio") new_gallery = gr.Gallery(label="New LoRAs", elem_id="gallery_new", columns=3, value=[(item["image"], item["title"]) for item in sdxl_loras_raw_new], allow_preview=False, show_share_button=False) gallery = gr.Gallery( #value=[(item["image"], item["title"]) for item in sdxl_loras], label="SDXL LoRA Gallery", allow_preview=False, columns=3, elem_id="gallery", show_share_button=False, height=784 ) with gr.Column(): prompt_title = gr.Markdown( value="### Click on a LoRA in the gallery to select it", visible=True, elem_id="selected_lora", ) with gr.Row(): prompt = gr.Textbox(label="Prompt", show_label=False, lines=1, max_lines=1, placeholder="Type a prompt after selecting a LoRA", elem_id="prompt") button = gr.Button("Run", elem_id="run_button") with gr.Group(elem_id="share-btn-container", visible=False) 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") result = gr.Image( interactive=False, label="Generated Image", elem_id="result-image" ) face_strength = gr.Slider(0, 1, value=0.85, step=0.01, label="Face strength", info="Higher values increase the face likeness but reduce the creative liberty of the models") image_strength = gr.Slider(0, 1, value=0.15, step=0.01, label="Image strength", info="Higher values increase the similarity with the structure/colors of the original photo") with gr.Accordion("Advanced options", open=False): negative = gr.Textbox(label="Negative Prompt") weight = gr.Slider(0, 10, value=0.9, step=0.1, label="LoRA weight") guidance_scale = gr.Slider(0, 50, value=7, step=0.1, label="Guidance Scale") depth_control_scale = gr.Slider(0, 1, value=0.8, step=0.01, label="Zoe Depth ControlNet strenght") with gr.Column(elem_id="extra_info"): with gr.Accordion( "Use it with: 🧨 diffusers, ComfyUI, Invoke AI, SD.Next, AUTO1111", open=False, elem_id="accordion", ): with gr.Row(): use_diffusers = gr.Markdown("""## Select a LoRA first 🤗""") use_uis = gr.Markdown() with gr.Accordion("Submit a LoRA! 📥", open=False): submit_title = gr.Markdown( "### Streamlined submission coming soon! Until then [suggest your LoRA in the community tab](https://huggingface.co/spaces/multimodalart/LoraTheExplorer/discussions) 🤗" ) with gr.Group(elem_id="soon"): submit_source = gr.Radio( ["Hugging Face", "CivitAI"], label="LoRA source", value="Hugging Face", ) with gr.Row(): submit_source_hf = gr.Textbox( label="Hugging Face Model Repo", info="In the format `username/model_id`", ) submit_safetensors_hf = gr.Textbox( label="Safetensors filename", info="The filename `*.safetensors` in the model repo", ) with gr.Row(): submit_trigger_word_hf = gr.Textbox(label="Trigger word") submit_image = gr.Image( label="Example image (optional if the repo already contains images)" ) submit_button = gr.Button("Submit!") submit_disclaimer = gr.Markdown( "This is a curated gallery by me, [apolinário (multimodal.art)](https://twitter.com/multimodalart). I'll try to include as many cool LoRAs as they are submitted! You can [duplicate this Space](https://huggingface.co/spaces/multimodalart/LoraTheExplorer?duplicate=true) to use it privately, and add your own LoRAs by editing `sdxl_loras.json` in the Files tab of your private space." ) order_gallery.change( fn=swap_gallery, inputs=[order_gallery, gr_sdxl_loras], outputs=[gallery, gr_sdxl_loras], queue=False ) gallery.select( fn=update_selection, inputs=[gr_sdxl_loras], outputs=[prompt_title, prompt, prompt, selected_state, use_diffusers, use_uis, new_gallery], queue=False, show_progress=False ) new_gallery.select( fn=update_selection, inputs=[gr_sdxl_loras_new, gr.State(True)], outputs=[prompt_title, prompt, prompt, selected_state, use_diffusers, use_uis, gallery], queue=False, show_progress=False ) prompt.submit( fn=check_selected, inputs=[selected_state], queue=False, show_progress=False ).success( fn=run_lora, inputs=[photo, prompt, negative, weight, selected_state, face_strength, image_strength, guidance_scale, depth_control_scale, gr_sdxl_loras, gr_sdxl_loras_new], outputs=[result, share_group], ) button.click( fn=check_selected, inputs=[selected_state], queue=False, show_progress=False ).success( fn=run_lora, inputs=[photo, prompt, negative, weight, selected_state, face_strength, image_strength, guidance_scale, depth_control_scale, gr_sdxl_loras, gr_sdxl_loras_new], outputs=[result, share_group], ) share_button.click(None, [], [], js=share_js) demo.load(fn=shuffle_gallery, inputs=[gr_sdxl_loras], outputs=[gallery, gr_sdxl_loras], queue=False, js=js) demo.queue(max_size=20) demo.launch(share=True)