import gradio as gr import random import os import io, base64 from PIL import Image import numpy import shortuuid latent = gr.Interface.load("spaces/multimodalart/latentdiffusion") rudalle = gr.Interface.load("spaces/multimodalart/rudalle") diffusion = gr.Interface.load("spaces/multimodalart/diffusion") print(diffusion) vqgan = gr.Interface.load("spaces/multimodalart/vqgan") def text2image_latent(text,steps,width,height,images,diversity): results = latent(text, steps, width, height, images, diversity) image_paths = [] image_arrays = [] for image in results[1]: image_str = image[0] image_str = image_str.replace("data:image/png;base64,","") decoded_bytes = base64.decodebytes(bytes(image_str, "utf-8")) img = Image.open(io.BytesIO(decoded_bytes)) url = shortuuid.uuid() temp_dir = './tmp' if not os.path.exists(temp_dir): os.makedirs(temp_dir, exist_ok=True) image_path = f'{temp_dir}/{url}.png' img.save(f'{temp_dir}/{url}.png') image_paths.append(image_path) return(image_paths) def text2image_rudalle(text,aspect,model): image = rudalle(text,aspect,model)[0] return([image]) def text2image_vqgan(text,width,height,style,steps,flavor): results = vqgan(text,width,height,style,steps,flavor) return([results]) def text2image_diffusion(text,steps_diff, images_diff, weight, clip): results = diffusion(text, steps_diff, images_diff, weight, clip) image_paths = [] print(results) for image in results: print('how many') image_str = image image_str = image_str.replace("data:image/png;base64,","") decoded_bytes = base64.decodebytes(bytes(image_str, "utf-8")) img = Image.open(io.BytesIO(decoded_bytes)) url = shortuuid.uuid() temp_dir = './tmp' if not os.path.exists(temp_dir): os.makedirs(temp_dir, exist_ok=True) image_path = f'{temp_dir}/{url}.png' img.save(f'{temp_dir}/{url}.png') image_paths.append(image_path) return(image_paths) css_mt = {"margin-top": "1em"} empty = gr.outputs.HTML() with gr.Blocks() as mindseye: gr.Markdown("
MindsEye Lite orchestrates multiple text-to-image models in one Spaces. This work carries the spirit of MindsEye Beta, but with simplified versions of the models due to current hardware limitations of Spaces. MindsEye Lite was created by @multimodalart, keep up with the latest multimodal ai art news here, join our Discord and consider supporting us on Patreon
") gr.Markdown("") text = gr.inputs.Textbox(placeholder="Type your prompt to generate an image", label="Prompt - try adding increments to your prompt such as 'a painting of', 'in the style of Picasso'", default="A giant mecha robot in Rio de Janeiro, oil on canvas") with gr.Row(): with gr.Column(): with gr.Tabs(): with gr.TabItem("Latent Diffusion"): gr.Markdown("Latent Diffusion is the state of the art of open source text-to-image models, superb in text synthesis. Sometimes struggles with complex prompts") steps = gr.inputs.Slider(label="Steps - more steps can increase quality but will take longer to generate",default=45,maximum=50,minimum=1,step=1) width = gr.inputs.Slider(label="Width", default=256, step=32, maximum=256, minimum=32) height = gr.inputs.Slider(label="Height", default=256, step=32, maximum = 256, minimum=32) images = gr.inputs.Slider(label="Images - How many images you wish to generate", default=2, step=1, minimum=1, maximum=4) diversity = gr.inputs.Slider(label="Diversity scale - How different from one another you wish the images to be",default=5.0, minimum=1.0, maximum=15.0) get_image_latent = gr.Button("Generate Image",css=css_mt) with gr.TabItem("VQGAN+CLIP"): gr.Markdown("VQGAN+CLIP is the most famous text-to-image generator. Can produce good artistic results") width_vq = gr.inputs.Slider(label="Width", default=256, minimum=32, step=32, maximum=512) height_vq= gr.inputs.Slider(label="Height", default=256, minimum=32, step=32, maximum=512) style = gr.inputs.Dropdown(label="Style - Hyper Fast Results is fast but compromises a bit of the quality",choices=["Default","Balanced","Detailed","Consistent Creativity","Realistic","Smooth","Subtle MSE","Hyper Fast Results"],default="Hyper Fast Results") steps_vq = gr.inputs.Slider(label="Steps - more steps can increase quality but will take longer to generate. All styles that are not Hyper Fast need at least 200 steps",default=50,maximum=300,minimum=1,step=1) flavor = gr.inputs.Dropdown(label="Flavor - pick a flavor for the style of the images, based on the images below",choices=["ginger", "cumin", "holywater", "zynth", "wyvern", "aaron", "moth", "juu"]) get_image_vqgan = gr.Button("Generate Image",css=css_mt) with gr.TabItem("Guided Diffusion"): gr.Markdown("Guided Diffusion models produce superb quality results. V-Diffusion is its latest implementation") steps_diff = gr.inputs.Slider(label="Steps - more steps can increase quality but will take longer to generate",default=40,maximum=80,minimum=1,step=1) images_diff = gr.inputs.Slider(label="Number of images in parallel", default=2, maximum=4, minimum=1, step=1) weight = gr.inputs.Slider(label="Weight - how closely the image should resemble the prompt", default=5, maximum=15, minimum=0, step=1) clip = gr.inputs.Checkbox(label="CLIP Guided - improves coherence with complex prompts, makes it slower") get_image_diffusion = gr.Button("Generate Image",css=css_mt) with gr.TabItem("ruDALLE"): gr.Markdown("ruDALLE is a replication of DALL-E 1 in the russian language. No worries, your prompts will be translated automatically to russian. In case you see an error, try again a few times") aspect = gr.inputs.Radio(label="Aspect Ratio", choices=["Square", "Horizontal", "Vertical"],default="Square") model = gr.inputs.Dropdown(label="Model", choices=["Surrealism","Realism", "Emoji"], default="Surrealism") get_image_rudalle = gr.Button("Generate Image",css=css_mt) with gr.Column(): with gr.Tabs(): #with gr.TabItem("Image output"): # image = gr.outputs.Image() with gr.TabItem("Gallery output"): gallery = gr.Gallery(label="Individual images") get_image_latent.click(text2image_latent, inputs=[text,steps,width,height,images,diversity], outputs=gallery) get_image_rudalle.click(text2image_rudalle, inputs=[text,aspect,model], outputs=gallery) get_image_vqgan.click(text2image_vqgan, inputs=[text,width_vq,height_vq,style,steps_vq,flavor],outputs=gallery) get_image_diffusion.click(text2image_diffusion, inputs=[text, steps_diff, images_diff, weight, clip],outputs=gallery) mindseye.launch()