from __future__ import annotations import math import random import spaces import gradio as gr import numpy as np import torch from PIL import Image from diffusers import EDMEulerScheduler, StableDiffusionXLInstructPix2PixPipeline from huggingface_hub import hf_hub_download, InferenceClient from diffusers import DiffusionPipeline dtype = torch.bfloat16 device = "cuda" if torch.cuda.is_available() else "cpu" pipe = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16, revision="refs/pr/1").to(device) help_text = """ To optimize image results: - Adjust the **Image CFG weight** if the image isn't changing enough or is changing too much. Lower it to allow bigger changes, or raise it to preserve original details. - Modify the **Text CFG weight** to influence how closely the edit follows text instructions. Increase it to adhere more to the text, or decrease it for subtler changes. - Experiment with different **random seeds** and **CFG values** for varied outcomes. - **Rephrase your instructions** for potentially better results. - **Increase the number of steps** for enhanced edits. """ def set_timesteps_patched(self, num_inference_steps: int, device = None): self.num_inference_steps = num_inference_steps ramp = np.linspace(0, 1, self.num_inference_steps) sigmas = torch.linspace(math.log(self.config.sigma_min), math.log(self.config.sigma_max), len(ramp)).exp().flip(0) sigmas = (sigmas).to(dtype=torch.float32, device=device) self.timesteps = self.precondition_noise(sigmas) self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) self._step_index = None self._begin_index = None self.sigmas = self.sigmas.to("cpu") # Image Editor edit_file = hf_hub_download(repo_id="stabilityai/cosxl", filename="cosxl_edit.safetensors") EDMEulerScheduler.set_timesteps = set_timesteps_patched pipe_edit = StableDiffusionXLInstructPix2PixPipeline.from_single_file( edit_file, num_in_channels=8, is_cosxl_edit=True, vae=vae, torch_dtype=torch.float16 ) pipe_edit.scheduler = EDMEulerScheduler(sigma_min=0.002, sigma_max=120.0, sigma_data=1.0, prediction_type="v_prediction") pipe_edit.to("cuda") client1 = InferenceClient("HuggingFaceH4/zephyr-7b-beta") system_instructions1 = "<|system|>\nAct as Image Prompt Generation expert, Your task is to modify prompt by USER to more better and detailed prompt for Image Generation. \n Ensure the prompt is deatiled, yet descriptive to generate an exceptional image that meets the user's expectations. \n Your task is to reply with final optimized prompt only. Reply with optimized prompt only.\n<|user|>\n" def promptifier(prompt): formatted_prompt = f"{system_instructions1}{prompt}\n<|assistant|>\n" stream = client1.text_generation(formatted_prompt, max_new_tokens=300) return stream # Generator @spaces.GPU(duration=60, queue=False) def king(type , input_image , instruction: str , negative_prompt: str ="", enhance_prompt: bool = True, steps: int = 25, randomize_seed: bool = True, seed: int = 2404, width: int = 1024, height: int = 1024, guidance_scale: float = 6, progress=gr.Progress(track_tqdm=True) ): if type=="Image Editing" : input_image = Image.open(input_image).convert('RGB') if randomize_seed: seed = random.randint(0, 999999) generator = torch.manual_seed(seed) output_image = pipe_edit( instruction, negative_prompt=negative_prompt, image=input_image, guidance_scale=guidance_scale, image_guidance_scale=1.5, width = input_image.width, height = input_image.height, num_inference_steps=steps, generator=generator, output_type="latent", ).images refine = refiner( prompt=f"{instruction}, 4k, hd, high quality, masterpiece", negative_prompt = negative_prompt, guidance_scale=7.5, num_inference_steps=steps, image=output_image, generator=generator, ).images[0] return seed, refine else : if randomize_seed: seed = random.randint(0, 999999) generator = torch.Generator().manual_seed(seed) if enhance_prompt: print(f"BEFORE: {instruction} ") instruction = promptifier(instruction) print(f"AFTER: {instruction} ") image = pipe( prompt = instruction, negative_prompt = negative_prompt, width = width, height = height, num_inference_steps = (steps/5), generator = generator, guidance_scale=0.0, output_type="latent" ).images refine = refiner( prompt=instruction, negative_prompt = negative_prompt, guidance_scale = 7.5, num_inference_steps= steps, image=image, generator=generator, ).images[0] return seed, refine client = InferenceClient() # Prompt classifier def response(instruction, input_image=None ): if input_image is None: output="Image Generation" else: try: text = instruction labels = ["Image Editing", "Image Generation"] classification = client.zero_shot_classification(text, labels, multi_label=True) output = classification[0] output = str(output) if "Editing" in output: output = "Image Editing" else: output = "Image Generation" except: if input_image is None: output="Image Generation" else: output="Image Editing" return output css = ''' .gradio-container{max-width: 700px !important} h1{text-align:center} footer { visibility: hidden } ''' examples=[ [ "Image Generation", None, "A luxurious supercar with a unique design. The car should have a pearl white finish, and gold accents. 4k, realistic.", ], [ "Image Editing", "./supercar.png", "make it red", ], [ "Image Editing", "./red_car.png", "add some snow", ], [ "Image Generation", None, "An alien grasping a sign board contain word 'ALIEN' with Neon Glow, neon, futuristic, neonpunk, neon lights", ], [ "Image Generation", None, "Beautiful Eiffel Tower at Night", ], [ "Image Generation", None, "Beautiful Eiffel Tower at Night", ], ] with gr.Blocks(css=css) as demo: gr.Markdown("# Image Generation , Image Editing \n ### Note: First image generation takes time") with gr.Row(): instruction = gr.Textbox(lines=1, label="Instruction", interactive=True) generate_button = gr.Button("Run", scale=0) with gr.Row(): type = gr.Dropdown(["Image Generation","Image Editing"], label="Task", value="Image Generation",interactive=True) enhance_prompt = gr.Checkbox(label="Enhance prompt", value=False, scale=0) with gr.Row(): input_image = gr.Image(label="Image", type='filepath', interactive=True) with gr.Row(): guidance_scale = gr.Number(value=6.0, step=0.1, label="Guidance Scale", interactive=True) steps = gr.Number(value=25, step=1, label="Steps", interactive=True) with gr.Accordion("Advanced options", open=False): with gr.Row(): negative_prompt = gr.Text( label="Negative prompt", max_lines=1, value="(deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, ugly, disgusting, blurry, amputation,(face asymmetry, eyes asymmetry, deformed eyes, open mouth)", visible=True) with gr.Row(): width = gr.Slider( label="Width", minimum=256, maximum=2048, step=64, value=1024) height = gr.Slider( label="Height", minimum=256, maximum=2048, step=64, value=1024) with gr.Row(): randomize_seed = gr.Checkbox(label="Randomize Seed", value = True, interactive=True ) seed = gr.Number(value=2404, step=1, label="Seed", interactive=True) gr.Examples( examples=examples, inputs=[type,input_image, instruction], fn=king, outputs=[input_image], cache_examples=False, ) # gr.Markdown(help_text) instruction.change(fn=response, inputs=[instruction,input_image], outputs=type, queue=False) input_image.upload(fn=response, inputs=[instruction,input_image], outputs=type, queue=False) gr.on(triggers=[ generate_button.click, instruction.submit ], fn=king, inputs=[type, input_image, instruction, negative_prompt, enhance_prompt, steps, randomize_seed, seed, width, height, guidance_scale, ], outputs=[seed, input_image], api_name = "image_gen_pro", queue=False ) demo.queue(max_size=500).launch()