import PIL import requests import torch import gradio as gr import random from PIL import Image import os import time from diffusers import StableDiffusionInstructPix2PixPipeline, EulerAncestralDiscreteScheduler #Loading from Diffusers Library model_id = "timbrooks/instruct-pix2pix" pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(model_id, torch_dtype=torch.float16, revision="fp16", safety_checker=None) pipe.to("cuda") #pipe.enable_attention_slicing() pipe.enable_xformers_memory_efficient_attention() pipe.unet.to(memory_format=torch.channels_last) help_text = """ **Note: Please be advised that a safety checker has been implemented in this public space. Any attempts to generate inappropriate or NSFW images will result in the display of a black screen as a precautionary measure to protect all users. We appreciate your cooperation in maintaining a safe and appropriate environment for all members of our community.** New features and bug-fixes: 1. Chat style interface 2. Now use **'reverse'** as prompt to get back the previous image after an unwanted edit 3. Use **'restart'** as prompt to get back to original image and start over! 4. Now you can load larger image files (~5 mb) as well Some notes from the official [instruct-pix2pix](https://huggingface.co/spaces/timbrooks/instruct-pix2pix) Space by the authors and from the official [Diffusers docs](https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/pix2pix) - If you're not getting what you want, there may be a few reasons: 1. Is the image not changing enough? Your guidance_scale may be too low. It should be >1. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, usually at the expense of lower image quality. This value dictates how similar the output should be to the input. This pipeline requires a value of at least `1`. It's possible your edit requires larger changes from the original image. 2. Alternatively, you can toggle image_guidance_scale. Image guidance scale is to push the generated image towards the inital image. Image guidance scale is enabled by setting `image_guidance_scale > 1`. Higher image guidance scale encourages to generate images that are closely linked to the source image `image`, usually at the expense of lower image quality. 3. I have observed that rephrasing the instruction sometimes improves results (e.g., "turn him into a dog" vs. "make him a dog" vs. "as a dog"). 4. Increasing the number of steps sometimes improves results. 5. Do faces look weird? The Stable Diffusion autoencoder has a hard time with faces that are small in the image. Try: * Cropping the image so the face takes up a larger portion of the frame. """ def previous(image): return image def upload_image(file): return Image.open(file) def upload_button_config(): return gr.update(visible=False) def upload_textbox_config(text_in): return gr.update(visible=True) def dummy_fn(): return 'dummy' def chat(btn_upload, image_in, in_steps, in_guidance_scale, in_img_guidance_scale, image_hid, img_name, counter_out, image_oneup, prompt, history, progress=gr.Progress(track_tqdm=True)): progress(0, desc="Starting...") if prompt != '' and prompt.lower() == 'reverse' : #--to add revert functionality later history = history or [] temp_img_name = img_name[:-4]+str(int(time.time()))+'.png' image_oneup.save(temp_img_name) response = 'Reverted to the last image ' + '' history.append((prompt, response)) return history, history, image_oneup, temp_img_name, counter_out if prompt != '' and prompt.lower() == 'restart' : #--to add revert functionality later history = history or [] temp_img_name = img_name[:-4]+str(int(time.time()))+'.png' #Resizing the image basewidth = 512 wpercent = (basewidth/float(image_in.size[0])) hsize = int((float(image_in.size[1])*float(wpercent))) image_in = image_in.resize((basewidth,hsize), Image.Resampling.LANCZOS) image_in.save(temp_img_name) response = 'Reverted to the last image ' + '' history.append((prompt, response)) return history, history, image_in, temp_img_name, counter_out #adding supportive sample text add_text_list = ["There you go", "Enjoy your image!", "Nice work! Wonder what you gonna do next!", "Way to go!", "Does this work for you?", "Something like this?"] if counter_out == 0: t1 = time.time() print(f"Time at start = {t1}") seed = random.randint(0, 1000000) img_name = f"./edited_image_{seed}.png" #convert file object to image image_in = Image.open(btn_upload) #Resizing the image basewidth = 512 wpercent = (basewidth/float(image_in.size[0])) hsize = int((float(image_in.size[1])*float(wpercent))) image_in = image_in.resize((basewidth,hsize), Image.Resampling.LANCZOS) #if os.path.exists(img_name): # os.remove(img_name) #with open(img_name, "wb") as fp: # Save the image to the file-like object image_in.save(img_name) #Get the name of the saved image #saved_image_name0 = fp.name history = history or [] response = '' history.append((prompt, response)) counter_out += 1 t2 = time.time() print(f"Time at end = {t2}") time_diff = t2-t1 print(f"Time taken = {time_diff}") return history, history, image_in, img_name, counter_out elif counter_out == 1: #instruct-pix2pix inference edited_image = pipe(prompt, image=image_in, num_inference_steps=int(in_steps), guidance_scale=float(in_guidance_scale), image_guidance_scale=float(in_img_guidance_scale)).images[0] if os.path.exists(img_name): os.remove(img_name) temp_img_name = img_name[:-4]+str(int(time.time()))[-4:] +'.png' with open(temp_img_name, "wb") as fp: # Save the image to the file-like object edited_image.save(fp) #Get the name of the saved image saved_image_name1 = fp.name history = history or [] response = random.choice(add_text_list) + '' #IMG_NAME history.append((prompt, response)) counter_out += 1 return history, history, edited_image, temp_img_name, counter_out elif counter_out > 1: edited_image = pipe(prompt, image=image_hid, num_inference_steps=int(in_steps), guidance_scale=float(in_guidance_scale), image_guidance_scale=float(in_img_guidance_scale)).images[0] if os.path.exists(img_name): os.remove(img_name) temp_img_name = img_name[:-4]+str(int(time.time()))[-4:]+'.png' # Create a file-like object with open(temp_img_name, "wb") as fp: # Save the image to the file-like object edited_image.save(fp) #Get the name of the saved image saved_image_name2 = fp.name #edited_image.save(temp_img_name) #, overwrite=True) history = history or [] response = random.choice(add_text_list) + '' history.append((prompt, response)) counter_out += 1 return history, history, edited_image, temp_img_name, counter_out #Blocks layout with gr.Blocks(css="style.css") as demo: with gr.Column(elem_id="col-container") as main_col: gr.HTML("""
For faster inference without waiting in the queue, you may duplicate the space and upgrade to GPU in settings Diffusers implementation of instruct-pix2pix - InstructPix2Pix: Learning to Follow Image Editing Instructions!