import gradio as gr import torch from PIL import Image from torchvision import transforms from diffusers import StableDiffusionPipeline, StableDiffusionImageVariationPipeline, DiffusionPipeline import numpy as np import pandas as pd import math from transformers import CLIPTextModel, CLIPTokenizer import os # model_id = "stabilityai/stable-diffusion-2-1-base" # text_model_id = "CompVis/stable-diffusion-v-1-4-original" # text_model_id = "CompVis/stable-diffusion-v1-4" text_model_id = "runwayml/stable-diffusion-v1-5" # text_model_id = "stabilityai/stable-diffusion-2-1-base" model_id = "lambdalabs/sd-image-variations-diffusers" clip_model_id = "openai/clip-vit-large-patch14-336" max_tabs = 10 input_images = [None for i in range(max_tabs)] input_prompts = [None for i in range(max_tabs)] embedding_plots = [None for i in range(max_tabs)] embedding_powers = [1. for i in range(max_tabs)] # global embedding_base64s embedding_base64s = [None for i in range(max_tabs)] # embedding_base64s = gr.State(value=[None for i in range(max_tabs)]) def image_to_embedding(input_im): tform = transforms.Compose([ transforms.ToTensor(), transforms.Resize( (224, 224), interpolation=transforms.InterpolationMode.BICUBIC, antialias=False, ), transforms.Normalize( [0.48145466, 0.4578275, 0.40821073], [0.26862954, 0.26130258, 0.27577711]), ]) inp = tform(input_im).to(device) dtype = next(pipe.image_encoder.parameters()).dtype image = inp.tile(1, 1, 1, 1).to(device=device, dtype=dtype) image_embeddings = pipe.image_encoder(image).image_embeds image_embeddings = image_embeddings[0] # image_embeddings /= image_embeddings.norm(dim=-1, keepdim=True) image_embeddings_np = image_embeddings.cpu().detach().numpy() return image_embeddings_np def prompt_to_embedding(prompt): # inputs = processor(prompt, images=imgs, return_tensors="pt", padding=True) inputs = processor(prompt, return_tensors="pt", padding='max_length', max_length=77) # labels = torch.tensor(labels) # prompt_tokens = inputs.input_ids[0] prompt_tokens = inputs.input_ids # image = inputs.pixel_values with torch.no_grad(): prompt_embededdings = model.get_text_features(prompt_tokens.to(device)) # prompt_embededdings /= prompt_embededdings.norm(dim=-1, keepdim=True) prompt_embededdings = prompt_embededdings[0].cpu().detach().numpy() return prompt_embededdings def embedding_to_image(embeddings): size = math.ceil(math.sqrt(embeddings.shape[0])) image_embeddings_square = np.pad(embeddings, (0, size**2 - embeddings.shape[0]), 'constant') image_embeddings_square.resize(size,size) embedding_image = Image.fromarray(image_embeddings_square, mode="L") return embedding_image def embedding_to_base64(embeddings): import base64 # ensure float16 embeddings = embeddings.astype(np.float16) embeddings_b64 = base64.urlsafe_b64encode(embeddings).decode() return embeddings_b64 def base64_to_embedding(embeddings_b64): import base64 embeddings = base64.urlsafe_b64decode(embeddings_b64) embeddings = np.frombuffer(embeddings, dtype=np.float16) # embeddings = torch.tensor(embeddings) return embeddings def main( # input_im, embeddings, scale=3.0, n_samples=4, steps=25, seed=None ): if seed == None: seed = np.random.randint(2147483647) # if device contains cuda if device.type == 'cuda': generator = torch.Generator(device=device).manual_seed(int(seed)) else: generator = torch.Generator().manual_seed(int(seed)) # use cpu as does not work on mps embeddings = base64_to_embedding(embeddings) embeddings = torch.tensor(embeddings, dtype=torch_size).to(device) images_list = pipe( # inp.tile(n_samples, 1, 1, 1), # [embeddings * n_samples], embeddings, guidance_scale=scale, num_inference_steps=steps, generator=generator, ) images = [] for i, image in enumerate(images_list["images"]): images.append(image) # images.append(embedding_image) return images def on_image_load_update_embeddings(image_data): # image to embeddings if image_data is None: # embeddings = prompt_to_embedding('') # embeddings_b64 = embedding_to_base64(embeddings) # return gr.Text.update(embeddings_b64) return gr.Text.update('') embeddings = image_to_embedding(image_data) embeddings_b64 = embedding_to_base64(embeddings) return gr.Text.update(embeddings_b64) def on_prompt_change_update_embeddings(prompt): # prompt to embeddings if prompt is None or prompt == "": embeddings = prompt_to_embedding('') embeddings_b64 = embedding_to_base64(embeddings) return gr.Text.update(embedding_to_base64(embeddings)) embeddings = prompt_to_embedding(prompt) embeddings_b64 = embedding_to_base64(embeddings) return gr.Text.update(embeddings_b64) def update_average_embeddings(embedding_base64s_state, embedding_powers): final_embedding = None num_embeddings = 0 for i, embedding_base64 in enumerate(embedding_base64s_state): if embedding_base64 is None or embedding_base64 == "": continue embedding = base64_to_embedding(embedding_base64) embedding = embedding * embedding_powers[i] if final_embedding is None: final_embedding = embedding else: final_embedding = final_embedding + embedding num_embeddings += 1 if final_embedding is None: # embeddings = prompt_to_embedding('') # embeddings_b64 = embedding_to_base64(embeddings) # return gr.Text.update(embeddings_b64) return gr.Text.update('') # TODO toggle this to support average or sum final_embedding = final_embedding / num_embeddings embeddings_b64 = embedding_to_base64(final_embedding) return embeddings_b64 def on_power_change_update_average_embeddings(embedding_base64s_state, embedding_power_state, power, idx): embedding_power_state[idx] = power embeddings_b64 = update_average_embeddings(embedding_base64s_state, embedding_power_state) return gr.Text.update(embeddings_b64) def on_embeddings_changed_update_average_embeddings(embedding_base64s_state, embedding_power_state, embedding_base64, idx): embedding_base64s_state[idx] = embedding_base64 if embedding_base64 != '' else None embeddings_b64 = update_average_embeddings(embedding_base64s_state, embedding_power_state) return gr.Text.update(embeddings_b64) def on_embeddings_changed_update_plot(embeddings_b64): # plot new embeddings if embeddings_b64 is None or embeddings_b64 == "": data = pd.DataFrame({ 'embedding': [], 'index': []}) return gr.LinePlot.update(data, x="index", y="embedding", # color="country", title="Embeddings", # stroke_dash="cluster", # x_lim=[1950, 2010], tooltip=['index', 'embedding'], # stroke_dash_legend_title="Country Cluster", # height=300, width=0) embeddings = base64_to_embedding(embeddings_b64) data = pd.DataFrame({ 'embedding': embeddings, 'index': [n for n in range(len(embeddings))]}) return gr.LinePlot.update(data, x="index", y="embedding", # color="country", title="Embeddings", # stroke_dash="cluster", # x_lim=[1950, 2010], tooltip=['index', 'embedding'], # stroke_dash_legend_title="Country Cluster", # height=300, width=embeddings.shape[0]) def on_example_image_click_set_image(input_image, image_url): input_image.value = image_url device = torch.device("mps" if torch.backends.mps.is_available() else "cuda:0" if torch.cuda.is_available() else "cpu") torch_size = torch.float16 if device == ('cuda') else torch.float32 # torch_size = torch.float32 pipe = StableDiffusionPipeline.from_pretrained( model_id, custom_pipeline="pipeline.py", torch_dtype=torch_size, # , revision="fp16", requires_safety_checker = False, safety_checker=None, text_encoder = CLIPTextModel, tokenizer = CLIPTokenizer, ) pipe = pipe.to(device) from transformers import AutoProcessor, AutoModel processor = AutoProcessor.from_pretrained(clip_model_id) model = AutoModel.from_pretrained(clip_model_id) model = model.to(device) examples = [ ["SohoJoeEth.jpeg", "Ray-Liotta-Goodfellas.jpg", "SohoJoeEth + Ray.jpeg"], # ["SohoJoeEth.jpeg", "Donkey.jpg", "SohoJoeEth + Donkey.jpeg"], # ["SohoJoeEth.jpeg", "Snoop Dogg.jpg", "SohoJoeEth + Snoop Dogg.jpeg"], ] tile_size = 100 # image_folder = os.path.join("file", "images") image_folder ="images" # image_examples = { # "452650": "452650.jpeg", # "Prompt 1": "a college dorm with a desk and bunk beds", # "371739": "371739.jpeg", # "Prompt 2": "a large banana is placed before a stuffed monkey.", # "557922": "557922.jpeg", # "Prompt 3": "a person sitting on a bench using a cell phone", # } tabbed_examples = { "CoCo": { "452650": "452650.jpeg", "Prompt 1": "a college dorm with a desk and bunk beds", "371739": "371739.jpeg", "Prompt 2": "a large banana is placed before a stuffed monkey.", "557922": "557922.jpeg", "Prompt 3": "a person sitting on a bench using a cell phone", "540554": "540554.jpeg", "Prompt 4": "two trains are coming down the tracks, a steam engine and a modern train.", }, "Transforms": { "ColorWheel001": "ColorWheel001.jpg", "ColorWheel001 BW": "ColorWheel001 BW.jpg", "ColorWheel002": "ColorWheel002.jpg", "ColorWheel002 BW": "ColorWheel002 BW.jpg", }, "Portraits": { "Snoop": "Snoop Dogg.jpg", "Snoop Prompt": "Snoop Dogg", "Ray": "Ray-Liotta-Goodfellas.jpg", "Ray Prompt": "Ray Liotta, Goodfellas", "Anya": "Anya Taylor-Joy 003.jpg", "Anya Prompt": "Anya Taylor-Joy, The Queen's Gambit", "Billie": "billie eilish 004.jpeg", "Billie Prompt": "Billie Eilish, blonde hair", "Lizzo": "Lizzo 001.jpeg", "Lizzo Prompt": "Lizzo,", "Donkey": "Donkey.jpg", "Donkey Prompt": "Donkey, from Shrek", }, "NFT's": { "SohoJoe": "SohoJoeEth.jpeg", "SohoJoe Prompt": "SohoJoe.Eth", "Mirai": "Mirai.jpg", "Mirai Prompt": "Mirai from White Rabbit, @shibuyaxyz", "OnChainMonkey": "OnChainMonkey-2278.jpg", "OCM Prompt": "On Chain Monkey", "Wassie": "Wassie 4498.jpeg", "Wassie Prompt": "Wassie by Wassies", }, "Pups": { "Pup1": "pup1.jpg", "Prompt": "Teacup Yorkies", "Pup2": "pup2.jpg", "Pup3": "pup3.jpg", "Pup4": "pup4.jpeg", "Pup5": "pup5.jpg", }, } image_examples_tile_size = 50 with gr.Blocks() as demo: with gr.Row(): with gr.Column(scale=5): gr.Markdown( """ # Soho-Clip A tool for exploring CLIP embedding spaces. Try uploading a few images and/or add some text prompts and click generate images. """) with gr.Column(scale=2, min_width=(tile_size+20)*3): with gr.Row(): with gr.Column(scale=1, min_width=tile_size): gr.Markdown("## Input 1") with gr.Column(scale=1, min_width=tile_size): gr.Markdown("## Input 2") with gr.Column(scale=1, min_width=tile_size): gr.Markdown("## Generates:") for example in examples: with gr.Row(): for example in example: with gr.Column(scale=1, min_width=tile_size): local_path = os.path.join(image_folder, example) gr.Image( value = local_path, shape=(tile_size,tile_size), show_label=False, interactive=False) \ .style(height=tile_size, width=tile_size) with gr.Row(): for i in range(max_tabs): with gr.Tab(f"Input {i+1}"): with gr.Row(): with gr.Column(scale=1, min_width=240): input_images[i] = gr.Image(label="Image Prompt", show_label=True) with gr.Column(scale=3, min_width=600): embedding_plots[i] = gr.LinePlot(show_label=False).style(container=False) # input_image.change(on_image_load, inputs= [input_image, plot]) with gr.Row(): with gr.Column(scale=2, min_width=240): input_prompts[i] = gr.Textbox(label="Text Prompt", show_label=True) with gr.Column(scale=3, min_width=600): with gr.Row(): # with gr.Slider(min=-5, max=5, value=1, label="Power", show_label=True): # embedding_powers[i] = gr.Slider.value embedding_powers[i] = gr.Slider(minimum=-3, maximum=3, value=1, label="Power", show_label=True, interactive=True) with gr.Row(): with gr.Accordion(f"Embeddings (base64)", open=False): embedding_base64s[i] = gr.Textbox(show_label=False) for idx, (tab_title, examples) in enumerate(tabbed_examples.items()): with gr.Tab(tab_title): with gr.Row(): for idx, (title, example) in enumerate(examples.items()): if example.endswith(".jpg") or example.endswith(".jpeg"): # add image example local_path = os.path.join(image_folder, example) with gr.Column(scale=1, min_width=image_examples_tile_size): gr.Examples( examples=[local_path], inputs=input_images[i], label=title, ) else: # add text example with gr.Column(scale=1, min_width=image_examples_tile_size*2): gr.Examples( examples=[example], inputs=input_prompts[i], label=title, ) with gr.Row(): average_embedding_plot = gr.LinePlot(show_label=True, label="Average Embeddings (base64)").style(container=False) with gr.Row(): with gr.Accordion(f"Avergage embeddings in base 64", open=False): average_embedding_base64 = gr.Textbox(show_label=False) with gr.Row(): submit = gr.Button("Generate images") with gr.Row(): with gr.Column(scale=1, min_width=200): scale = gr.Slider(0, 25, value=3, step=1, label="Guidance scale") with gr.Column(scale=1, min_width=200): n_samples = gr.Slider(1, 4, value=1, step=1, label="Number images") with gr.Column(scale=1, min_width=200): steps = gr.Slider(5, 50, value=25, step=5, label="Steps") with gr.Column(scale=1, min_width=200): seed = gr.Number(None, label="Seed (blank = random)", precision=0) with gr.Row(): output = gr.Gallery(label="Generated variations") embedding_base64s_state = gr.State(value=[None for i in range(max_tabs)]) embedding_power_state = gr.State(value=[1. for i in range(max_tabs)]) for i in range(max_tabs): input_images[i].change(on_image_load_update_embeddings, input_images[i], [embedding_base64s[i]]) input_prompts[i].change(on_prompt_change_update_embeddings, input_prompts[i], [embedding_base64s[i]]) embedding_base64s[i].change(on_embeddings_changed_update_plot, embedding_base64s[i], [embedding_plots[i]]) idx_state = gr.State(value=i) embedding_base64s[i].change(on_embeddings_changed_update_average_embeddings, [embedding_base64s_state, embedding_power_state, embedding_base64s[i], idx_state], average_embedding_base64) embedding_powers[i].change(on_power_change_update_average_embeddings, [embedding_base64s_state, embedding_power_state, embedding_powers[i], idx_state], average_embedding_base64) average_embedding_base64.change(on_embeddings_changed_update_plot, average_embedding_base64, average_embedding_plot) # submit.click(main, inputs= [embedding_base64s[0], scale, n_samples, steps, seed], outputs=output) submit.click(main, inputs= [average_embedding_base64, scale, n_samples, steps, seed], outputs=output) output.style(grid=2) with gr.Row(): gr.Markdown( """ My interest is to use CLIP for image/video understanding (see [CLIP_visual-spatial-reasoning](https://github.com/Sohojoe/CLIP_visual-spatial-reasoning).) ### Initial Features - Combine up to 10 Images and/or text inputs to create an average embedding space. - View embedding spaces as graph - Generate a new image based on the average embedding space ### Known limitations - Text input is a little off (requires fine tuning and I'm having issues with that at the moment) - It can only generate a single image at a time - Not easy to use the sample images ### Acknowledgements - I heavily build on Justin Pinkney's [Experiments in Image Variation](https://www.justinpinkney.com/image-variation-experiments). Please credit them if you use this work. - [CLIP](https://openai.com/blog/clip/) - [Stable Diffusion](https://github.com/CompVis/stable-diffusion) """) # ![Alt Text](file/pup1.jpg) # # ![Alt Text](file/pup1.jpg){height=100 width=100} if __name__ == "__main__": demo.launch()