import gradio as gr from diffusers import StableDiffusionXLPipeline, EDMEulerScheduler, StableDiffusionXLInstructPix2PixPipeline from custom_pipeline import CosStableDiffusionXLInstructPix2PixPipeline from huggingface_hub import hf_hub_download import numpy as np import math import spaces import torch from PIL import Image edit_file = hf_hub_download(repo_id="stabilityai/cosxl", filename="cosxl_edit.safetensors") normal_file = hf_hub_download(repo_id="stabilityai/cosxl", filename="cosxl.safetensors") 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") # to avoid too much CPU/GPU communication def resize_image(image, resolution): original_width, original_height = image.size if original_width > original_height: new_width = resolution new_height = int((resolution / original_width) * original_height) else: new_height = resolution new_width = int((resolution / original_height) * original_width) resized_img = image.resize((new_width, new_height), Image.Resampling.LANCZOS) return resized_img EDMEulerScheduler.set_timesteps = set_timesteps_patched pipe_edit = StableDiffusionXLInstructPix2PixPipeline.from_single_file( edit_file, num_in_channels=8, is_cosxl_edit=True, 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") pipe_normal = StableDiffusionXLPipeline.from_single_file(normal_file, torch_dtype=torch.float16) pipe_normal.scheduler = EDMEulerScheduler(sigma_min=0.002, sigma_max=120.0, sigma_data=1.0, prediction_type="v_prediction") pipe_normal.to("cuda") @spaces.GPU def run_normal(prompt, negative_prompt="", guidance_scale=7, steps=20, progress=gr.Progress(track_tqdm=True)): return pipe_normal(prompt, negative_prompt=negative_prompt, guidance_scale=guidance_scale, num_inference_steps=steps).images[0] @spaces.GPU def run_edit(image, prompt, negative_prompt="", guidance_scale=7, steps=20, progress=gr.Progress(track_tqdm=True)): image = resize_image(image, 1024) print("Image resized to ", image.size) width, height = image.size #image.resize((resolution, resolution)) return pipe_edit(prompt=prompt,image=image,height=height,width=width,negative_prompt=negative_prompt, guidance_scale=guidance_scale,num_inference_steps=steps).images[0] css = ''' .gradio-container{ max-width: 768px !important; margin: 0 auto; } ''' normal_examples = ["portrait photo of a girl, photograph, highly detailed face, depth of field, moody light, golden hour, style by Dan Winters, Russell James, Steve McCurry, centered, extremely detailed, Nikon D850, award winning photography", "backlit photography of a dog", "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k", "A photo of beautiful mountain with realistic sunset and blue lake, highly detailed, masterpiece"] edit_examples = [["mountain.png", "make it a cloudy day"], ["painting.png", "make the earring fancier"]] with gr.Blocks(css=css) as demo: gr.Markdown('''# CosXL demo Unofficial demo for CosXL, a SDXL model tuned to produce full color range images. CosXL Edit allows you to perform edits on images. Both have a [non-commercial community license](https://huggingface.co/stabilityai/cosxl/blob/main/LICENSE) ''') with gr.Tab("CosXL Edit"): with gr.Group(): image_edit = gr.Image(label="Image you would like to edit", type="pil") with gr.Row(): prompt_edit = gr.Textbox(show_label=False, scale=4, placeholder="Edit instructions, e.g.: Make the day cloudy") button_edit = gr.Button("Generate", min_width=120) output_edit = gr.Image(label="Your result image", interactive=False) with gr.Accordion("Advanced Settings", open=False): negative_prompt_edit = gr.Textbox(label="Negative Prompt") guidance_scale_edit = gr.Number(label="Guidance Scale", value=7) steps_edit = gr.Slider(label="Steps", minimum=10, maximum=50, value=20) gr.Examples(examples=edit_examples, fn=run_edit, inputs=[image_edit, prompt_edit], outputs=[output_edit], cache_examples=True) with gr.Tab("CosXL"): with gr.Group(): with gr.Row(): prompt_normal = gr.Textbox(show_label=False, scale=4, placeholder="Your prompt, e.g.: backlit photography of a dog") button_normal = gr.Button("Generate", min_width=120) output_normal = gr.Image(label="Your result image", interactive=False) with gr.Accordion("Advanced Settings", open=False): negative_prompt_normal = gr.Textbox(label="Negative Prompt") guidance_scale_normal = gr.Number(label="Guidance Scale", value=7) steps_normal = gr.Slider(label="Steps", minimum=10, maximum=50, value=20) gr.Examples(examples=normal_examples, fn=run_normal, inputs=[prompt_normal], outputs=[output_normal], cache_examples="lazy") gr.on( triggers=[ button_normal.click, prompt_normal.submit ], fn=run_normal, inputs=[prompt_normal, negative_prompt_normal, guidance_scale_normal, steps_normal], outputs=[output_normal], ) gr.on( triggers=[ button_edit.click, prompt_edit.submit ], fn=run_edit, inputs=[image_edit, prompt_edit, negative_prompt_edit, guidance_scale_edit, steps_edit], outputs=[output_edit] ) if __name__ == "__main__": demo.launch(share=True)