import os import spaces import torch from diffusers import StableDiffusionPipeline, EulerAncestralDiscreteScheduler import gradio as gr import random import tqdm from huggingface_hub import hf_hub_download from transformers import CLIPTextModel, CLIPTokenizer # Enable TQDM progress tracking tqdm.monitor_interval = 0 # Load the model from safetensors file def load_model(): model_path = hf_hub_download( repo_id="kayfahaarukku/AkashicPulse-v1.0", filename="AkashicPulse-v1.0-ft-ft.safetensors" ) # Initialize tokenizer and text encoder from standard SD 1.5 tokenizer = CLIPTokenizer.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="tokenizer") text_encoder = CLIPTextModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="text_encoder") # Initialize pipeline with text encoder and tokenizer pipe = StableDiffusionPipeline.from_single_file( model_path, torch_dtype=torch.float16, use_safetensors=True, tokenizer=tokenizer, text_encoder=text_encoder, requires_safety_checker=False, safety_checker=None ) pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config) return pipe # Load the pipeline pipe = load_model() # Function to generate an image @spaces.GPU def generate_image(prompt, negative_prompt, use_defaults, resolution, guidance_scale, num_inference_steps, seed, randomize_seed, progress=gr.Progress()): pipe.to('cuda') if randomize_seed: seed = random.randint(0, 99999999) if use_defaults: prompt = f"{prompt}, masterpiece, best quality" negative_prompt = f"lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, signature, watermark, username, blurry, {negative_prompt}" generator = torch.manual_seed(seed) def callback(step, timestep, latents): progress(step / num_inference_steps) return width, height = map(int, resolution.split('x')) # Add empty dict for additional kwargs added_cond_kwargs = {"text_embeds": None, "time_ids": None} image = pipe( prompt, negative_prompt=negative_prompt, width=width, height=height, guidance_scale=guidance_scale, num_inference_steps=num_inference_steps, generator=generator, callback=callback, callback_steps=1, added_cond_kwargs=added_cond_kwargs ).images[0] torch.cuda.empty_cache() metadata_text = f"{prompt}\nNegative prompt: {negative_prompt}\nSteps: {num_inference_steps}, Sampler: Euler a, Size: {width}x{height}, Seed: {seed}, CFG scale: {guidance_scale}" return image, seed, metadata_text # Define Gradio interface def interface_fn(prompt, negative_prompt, use_defaults, resolution, guidance_scale, num_inference_steps, seed, randomize_seed, progress=gr.Progress()): try: image, seed, metadata_text = generate_image(prompt, negative_prompt, use_defaults, resolution, guidance_scale, num_inference_steps, seed, randomize_seed, progress) return image, seed, gr.update(value=metadata_text) except Exception as e: print(f"Error generating image: {str(e)}") raise e def reset_inputs(): return gr.update(value=''), gr.update(value=''), gr.update(value=True), gr.update(value='832x1216'), gr.update(value=7), gr.update(value=28), gr.update(value=0), gr.update(value=True), gr.update(value='') with gr.Blocks(title="AkashicPulse v1.0 Demo", theme="NoCrypt/miku@1.2.1") as demo: gr.HTML( "

AkashicPulse v1.0 Demo

" "This demo showcases the AkashicPulse v1.0 model capabilities. For best results, it's recommended to run the model in Stable Diffusion WebUI or ComfyUI with MaHiRo CFG enabled." ) with gr.Row(): with gr.Column(): prompt_input = gr.Textbox(lines=2, placeholder="Enter prompt here", label="Prompt") negative_prompt_input = gr.Textbox(lines=2, placeholder="Enter negative prompt here", label="Negative Prompt") use_defaults_input = gr.Checkbox(label="Use Default Quality Tags and Negative Prompt", value=True) resolution_input = gr.Radio( choices=[ "1024x1024", "1152x896", "896x1152", "1216x832", "832x1216", "1344x768", "768x1344", "1536x640", "640x1536" ], label="Resolution", value="832x1216" ) guidance_scale_input = gr.Slider(minimum=4, maximum=10, step=0.5, label="Guidance Scale (CFG)", value=7) num_inference_steps_input = gr.Slider(minimum=20, maximum=30, step=1, label="Number of Steps", value=28) seed_input = gr.Slider(minimum=0, maximum=999999999, step=1, label="Seed", value=0, interactive=True) randomize_seed_input = gr.Checkbox(label="Randomize Seed", value=True) generate_button = gr.Button("Generate") reset_button = gr.Button("Reset") with gr.Column(): output_image = gr.Image(type="pil", label="Generated Image") with gr.Accordion("Parameters", open=False): gr.Markdown( """ This parameter is compatible with Stable Diffusion WebUI's parameter importer. """ ) metadata_textbox = gr.Textbox(lines=6, label="Image Parameters", interactive=False, max_lines=6) gr.Markdown( """ ### Recommended prompt formatting: `1girl/1boy, character name, series, by artist name, the rest of the prompt, masterpiece, best quality` **PS:** `masterpiece, best quality` is automatically added when "Use Default Quality Tags and Negative Prompt" is enabled ### Current settings (recommended): - Sampler: Euler a (fixed) - Steps: 20-30 (sweet spot: 28) - CFG: 4-10 (sweet spot: 7) - Optional: Enable MaHiRo CFG in reForge or ComfyUI """ ) generate_button.click( interface_fn, inputs=[ prompt_input, negative_prompt_input, use_defaults_input, resolution_input, guidance_scale_input, num_inference_steps_input, seed_input, randomize_seed_input ], outputs=[output_image, seed_input, metadata_textbox] ) reset_button.click( reset_inputs, inputs=[], outputs=[ prompt_input, negative_prompt_input, use_defaults_input, resolution_input, guidance_scale_input, num_inference_steps_input, seed_input, randomize_seed_input, metadata_textbox ] ) demo.queue(max_size=20).launch(share=False)