import torch import os import gradio as gr from PIL import Image import random from diffusers import ( DiffusionPipeline, AutoencoderKL, StableDiffusionControlNetPipeline, ControlNetModel, StableDiffusionLatentUpscalePipeline, StableDiffusionImg2ImgPipeline, StableDiffusionControlNetImg2ImgPipeline, DPMSolverMultistepScheduler, # <-- Added import EulerDiscreteScheduler # <-- Added import ) from share_btn import community_icon_html, loading_icon_html, share_js from gallery_history import fetch_gallery_history, show_gallery_history from illusion_style import css BASE_MODEL = "HusseinHE/icbinh" # Initialize both pipelines vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16) #init_pipe = DiffusionPipeline.from_pretrained("SG161222/Realistic_Vision_V5.1_noVAE", torch_dtype=torch.float16) controlnet = ControlNetModel.from_pretrained("monster-labs/control_v1p_sd15_qrcode_monster", torch_dtype=torch.float16)#, torch_dtype=torch.float16) main_pipe = StableDiffusionControlNetPipeline.from_pretrained( BASE_MODEL, controlnet=controlnet, vae=vae, safety_checker=None, torch_dtype=torch.float16, ).to("cuda") #main_pipe.unet = torch.compile(main_pipe.unet, mode="reduce-overhead", fullgraph=True) #main_pipe.unet.to(memory_format=torch.channels_last) #main_pipe.unet = torch.compile(main_pipe.unet, mode="reduce-overhead", fullgraph=True) #model_id = "stabilityai/sd-x2-latent-upscaler" image_pipe = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(BASE_MODEL, unet=main_pipe.unet, vae=vae, controlnet=controlnet, safety_checker=None, torch_dtype=torch.float16).to("cuda") #image_pipe.unet = torch.compile(image_pipe.unet, mode="reduce-overhead", fullgraph=True) #upscaler = StableDiffusionLatentUpscalePipeline.from_pretrained(model_id, torch_dtype=torch.float16) #upscaler.to("cuda") # Sampler map SAMPLER_MAP = { "DPM++ Karras SDE": lambda config: DPMSolverMultistepScheduler.from_config(config, use_karras=True, algorithm_type="sde-dpmsolver++"), "Euler": lambda config: EulerDiscreteScheduler.from_config(config), } def center_crop_resize(img, output_size=(512, 512)): width, height = img.size # Calculate dimensions to crop to the center new_dimension = min(width, height) left = (width - new_dimension)/2 top = (height - new_dimension)/2 right = (width + new_dimension)/2 bottom = (height + new_dimension)/2 # Crop and resize img = img.crop((left, top, right, bottom)) img = img.resize(output_size) return img def common_upscale(samples, width, height, upscale_method, crop=False): if crop == "center": old_width = samples.shape[3] old_height = samples.shape[2] old_aspect = old_width / old_height new_aspect = width / height x = 0 y = 0 if old_aspect > new_aspect: x = round((old_width - old_width * (new_aspect / old_aspect)) / 2) elif old_aspect < new_aspect: y = round((old_height - old_height * (old_aspect / new_aspect)) / 2) s = samples[:,:,y:old_height-y,x:old_width-x] else: s = samples return torch.nn.functional.interpolate(s, size=(height, width), mode=upscale_method) def upscale(samples, upscale_method, scale_by): #s = samples.copy() width = round(samples["images"].shape[3] * scale_by) height = round(samples["images"].shape[2] * scale_by) s = common_upscale(samples["images"], width, height, upscale_method, "disabled") return (s) # Inference function def inference( control_image: Image.Image, prompt: str, negative_prompt: str, guidance_scale: float = 8.0, controlnet_conditioning_scale: float = 1, control_guidance_start: float = 1, control_guidance_end: float = 1, upscaler_strength: float = 0.5, seed: int = -1, sampler = "DPM++ Karras SDE", progress = gr.Progress(track_tqdm=True) ): if prompt is None or prompt == "": raise gr.Error("Prompt is required") # Generate the initial image #init_image = init_pipe(prompt).images[0] # Rest of your existing code control_image_small = center_crop_resize(control_image) main_pipe.scheduler = SAMPLER_MAP[sampler](main_pipe.scheduler.config) my_seed = random.randint(0, 2**32 - 1) if seed == -1 else seed generator = torch.manual_seed(my_seed) out = main_pipe( prompt=prompt, negative_prompt=negative_prompt, image=control_image_small, guidance_scale=float(guidance_scale), controlnet_conditioning_scale=float(controlnet_conditioning_scale), generator=generator, control_guidance_start=float(control_guidance_start), control_guidance_end=float(control_guidance_end), num_inference_steps=30, output_type="latent" ) control_image_large = center_crop_resize(control_image, (1024, 1024)) upscaled_latents = upscale(out, "nearest-exact", 2) out_image = image_pipe( prompt=prompt, negative_prompt=negative_prompt, control_image=control_image_large, image=upscaled_latents, guidance_scale=float(guidance_scale), generator=generator, num_inference_steps=30, strength=upscaler_strength, control_guidance_start=float(control_guidance_start), control_guidance_end=float(control_guidance_end), controlnet_conditioning_scale=float(controlnet_conditioning_scale) ) return out_image["images"][0], gr.update(visible=True), my_seed #return out with gr.Blocks(css=css) as app: gr.Markdown( '''