import spaces import gradio as gr import numpy as np import random import torch from diffusers import StableDiffusion3Pipeline, AutoencoderKL from transformers import CLIPTextModelWithProjection, T5EncoderModel from transformers import CLIPTokenizer, T5TokenizerFast import re import paramiko import urllib import time import os from image_gen_aux import UpscaleWithModel from huggingface_hub import hf_hub_download import datetime import cyper #from diffusers import SD3Transformer2DModel, AutoencoderKL #from models.transformer_sd3 import SD3Transformer2DModel #from pipeline_stable_diffusion_3_ipa import StableDiffusion3Pipeline from PIL import Image torch.backends.cuda.matmul.allow_tf32 = False torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False torch.backends.cudnn.allow_tf32 = False torch.backends.cudnn.deterministic = False torch.backends.cudnn.benchmark = False #torch.backends.cuda.preferred_blas_library="cublas" #torch.backends.cuda.preferred_linalg_library="cusolver" hftoken = os.getenv("HF_AUTH_TOKEN") code = r''' import torch import paramiko import os FTP_HOST = os.getenv("FTP_HOST") FTP_USER = os.getenv("FTP_USER") FTP_PASS = os.getenv("FTP_PASS") FTP_DIR = os.getenv("FTP_DIR") def upload_to_ftp(filename): try: transport = paramiko.Transport((FTP_HOST, 22)) destination_path=FTP_DIR+filename transport.connect(username = FTP_USER, password = FTP_PASS) sftp = paramiko.SFTPClient.from_transport(transport) sftp.put(filename, destination_path) sftp.close() transport.close() print(f"Uploaded {filename} to FTP server") except Exception as e: print(f"FTP upload error: {e}") ''' pyx = cyper.inline(code, fast_indexing=True, directives=dict(boundscheck=False, wraparound=False, language_level=3)) device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") vaeX=AutoencoderKL.from_pretrained("ford442/stable-diffusion-3.5-large-fp32", safety_checker=None, use_safetensors=True, subfolder='vae', low_cpu_mem_usage=False, token=True) pipe = StableDiffusion3Pipeline.from_pretrained( #"stabilityai # stable-diffusion-3.5-large", "ford442/stable-diffusion-3.5-large-bf16", # vae=AutoencoderKL.from_pretrained("ford442/stable-diffusion-3.5-large-fp32", use_safetensors=True, subfolder='vae',token=True), #scheduler = FlowMatchHeunDiscreteScheduler.from_pretrained('ford442/stable-diffusion-3.5-large-bf16', subfolder='scheduler',token=True), text_encoder=None, #CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder', token=True), # text_encoder=CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder', token=True), text_encoder_2=None, #CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_2',token=True), # text_encoder_2=CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_2',token=True), text_encoder_3=None, #T5EncoderModel.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_3',token=True), # text_encoder_3=T5EncoderModel.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_3',token=True), #tokenizer=CLIPTokenizer.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", add_prefix_space=True, subfolder="tokenizer", token=True), #tokenizer_2=CLIPTokenizer.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", add_prefix_space=True, subfolder="tokenizer_2", token=True), tokenizer_3=T5TokenizerFast.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", add_prefix_space=False, use_fast=True, subfolder="tokenizer_3", token=True), vae=None, #torch_dtype=torch.bfloat16, #use_safetensors=False, ) #pipe.load_lora_weights("ford442/sdxl-vae-bf16", weight_name="LoRA/UltraReal.safetensors") pipe.to(device=device, dtype=torch.bfloat16) #pipe.to(device) pipe.vae=vaeX.to('cpu') text_encoder=CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder', token=True).to(device=device, dtype=torch.bfloat16) text_encoder_2=CLIPTextModelWithProjection.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_2',token=True).to(device=device, dtype=torch.bfloat16) text_encoder_3=T5EncoderModel.from_pretrained("ford442/stable-diffusion-3.5-large-bf16", subfolder='text_encoder_3',token=True).to(device=device, dtype=torch.bfloat16) upscaler_2 = UpscaleWithModel.from_pretrained("Kim2091/ClearRealityV1").to('cpu') #.to(device) MAX_SEED = np.iinfo(np.int32).max MAX_IMAGE_SIZE = 4096 @spaces.GPU(duration=40) def infer_30( prompt, negative_prompt_1, negative_prompt_2, negative_prompt_3, width, height, guidance_scale, num_inference_steps, progress=gr.Progress(track_tqdm=True), ): pipe.text_encoder=text_encoder pipe.text_encoder_2=text_encoder_2 pipe.text_encoder_3=text_encoder_3 torch.set_float32_matmul_precision("highest") seed = random.randint(0, MAX_SEED) generator = torch.Generator(device='cuda').manual_seed(seed) print('-- generating image --') sd_image = pipe( prompt=prompt, prompt_2=prompt, prompt_3=prompt, negative_prompt=negative_prompt_1, negative_prompt_2=negative_prompt_2, negative_prompt_3=negative_prompt_3, guidance_scale=guidance_scale, num_inference_steps=num_inference_steps, width=width, height=height, # cross_attention_kwargs={"scale": 0.75}, generator=generator, max_sequence_length=512 ).images[0] print('-- got image --') timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S") sd35_path = f"sd35l_{timestamp}.png" sd_image.save(sd35_path,optimize=False,compress_level=0) pyx.upload_to_ftp(sd35_path) # pipe.unet.to('cpu') upscaler_2.to(torch.device('cuda')) with torch.no_grad(): upscale2 = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256) print('-- got upscaled image --') downscale2 = upscale2.resize((upscale2.width // 4, upscale2.height // 4),Image.LANCZOS) upscale_path = f"sd35l_upscale_{timestamp}.png" downscale2.save(upscale_path,optimize=False,compress_level=0) pyx.upload_to_ftp(upscale_path) return sd_image, prompt @spaces.GPU(duration=70) def infer_60( prompt, negative_prompt_1, negative_prompt_2, negative_prompt_3, width, height, guidance_scale, num_inference_steps, progress=gr.Progress(track_tqdm=True), ): pipe.text_encoder=text_encoder pipe.text_encoder_2=text_encoder_2 pipe.text_encoder_3=text_encoder_3 torch.set_float32_matmul_precision("highest") seed = random.randint(0, MAX_SEED) generator = torch.Generator(device='cuda').manual_seed(seed) print('-- generating image --') sd_image = pipe( prompt=prompt, prompt_2=prompt, prompt_3=prompt, negative_prompt=negative_prompt_1, negative_prompt_2=negative_prompt_2, negative_prompt_3=negative_prompt_3, guidance_scale=guidance_scale, num_inference_steps=num_inference_steps, width=width, height=height, generator=generator, max_sequence_length=512 ).images[0] print('-- got image --') timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S") sd35_path = f"sd35l_{timestamp}.png" sd_image.save(sd35_path,optimize=False,compress_level=0) pyx.upload_to_ftp(sd35_path) # pipe.unet.to('cpu') upscaler_2.to(torch.device('cuda')) with torch.no_grad(): upscale2 = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256) print('-- got upscaled image --') downscale2 = upscale2.resize((upscale2.width // 4, upscale2.height // 4),Image.LANCZOS) upscale_path = f"sd35l_upscale_{timestamp}.png" downscale2.save(upscale_path,optimize=False,compress_level=0) pyx.upload_to_ftp(upscale_path) return sd_image, prompt @spaces.GPU(duration=100) def infer_90( prompt, negative_prompt_1, negative_prompt_2, negative_prompt_3, width, height, guidance_scale, num_inference_steps, progress=gr.Progress(track_tqdm=True), ): pipe.text_encoder=text_encoder pipe.text_encoder_2=text_encoder_2 pipe.text_encoder_3=text_encoder_3 torch.set_float32_matmul_precision("highest") seed = random.randint(0, MAX_SEED) generator = torch.Generator(device='cuda').manual_seed(seed) print('-- generating image --') sd_image = pipe( prompt=prompt, prompt_2=prompt, prompt_3=prompt, negative_prompt=negative_prompt_1, negative_prompt_2=negative_prompt_2, negative_prompt_3=negative_prompt_3, guidance_scale=guidance_scale, num_inference_steps=num_inference_steps, width=width, height=height, generator=generator, max_sequence_length=512 ).images[0] print('-- got image --') timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S") sd35_path = f"sd35l_{timestamp}.png" sd_image.save(sd35_path,optimize=False,compress_level=0) pyx.upload_to_ftp(sd35_path) # pipe.unet.to('cpu') upscaler_2.to(torch.device('cuda')) with torch.no_grad(): upscale2 = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256) print('-- got upscaled image --') downscale2 = upscale2.resize((upscale2.width // 4, upscale2.height // 4),Image.LANCZOS) upscale_path = f"sd35l_upscale_{timestamp}.png" downscale2.save(upscale_path,optimize=False,compress_level=0) pyx.upload_to_ftp(upscale_path) return sd_image, prompt @spaces.GPU(duration=110) def infer_100( prompt, negative_prompt_1, negative_prompt_2, negative_prompt_3, width, height, guidance_scale, num_inference_steps, progress=gr.Progress(track_tqdm=True), ): torch.set_float32_matmul_precision("highest") seed = random.randint(0, MAX_SEED) generator = torch.Generator(device='cuda').manual_seed(seed) print('-- generating image --') sd_image = pipe( prompt=prompt, prompt_2=prompt, prompt_3=prompt, negative_prompt=negative_prompt_1, negative_prompt_2=negative_prompt_2, negative_prompt_3=negative_prompt_3, guidance_scale=guidance_scale, num_inference_steps=num_inference_steps, width=width, height=height, generator=generator, max_sequence_length=512 ).images[0] print('-- got image --') timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S") sd35_path = f"sd35l_{timestamp}.png" sd_image.save(sd35_path,optimize=False,compress_level=0) pyx.upload_to_ftp(sd35_path) # pipe.unet.to('cpu') upscaler_2.to(torch.device('cuda')) with torch.no_grad(): upscale2 = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256) print('-- got upscaled image --') downscale2 = upscale2.resize((upscale2.width // 4, upscale2.height // 4),Image.LANCZOS) upscale_path = f"sd35l_upscale_{timestamp}.png" downscale2.save(upscale_path,optimize=False,compress_level=0) pyx.upload_to_ftp(upscale_path) return sd_image, prompt css = """ #col-container {margin: 0 auto;max-width: 640px;} body{background-color: blue;} """ with gr.Blocks(theme=gr.themes.Origin(),css=css) as demo: with gr.Column(elem_id="col-container"): gr.Markdown(" # Text-to-Image StableDiffusion 3.5 Large") expanded_prompt_output = gr.Textbox(label="Prompt", lines=1) # Add this line with gr.Row(): prompt = gr.Text( label="Prompt", show_label=False, max_lines=1, placeholder="Enter your prompt", container=False, ) run_button_30 = gr.Button("Run 30", scale=0, variant="primary") run_button_60 = gr.Button("Run 60", scale=0, variant="primary") run_button_90 = gr.Button("Run 90", scale=0, variant="primary") run_button_100 = gr.Button("Run 100", scale=0, variant="primary") result = gr.Image(label="Result", show_label=False) with gr.Accordion("Advanced Settings", open=True): negative_prompt_1 = gr.Text( label="Negative prompt 1", max_lines=1, placeholder="Enter a negative prompt", visible=True, value="bad anatomy, poorly drawn hands, distorted face, blurry, out of frame, low resolution, grainy, pixelated, disfigured, mutated, extra limbs, bad composition" ) negative_prompt_2 = gr.Text( label="Negative prompt 2", max_lines=1, placeholder="Enter a second negative prompt", visible=True, value="unrealistic, cartoon, anime, sketch, painting, drawing, illustration, graphic, digital art, render, 3d, blurry, deformed, disfigured, poorly drawn, bad anatomy, mutated, extra limbs, ugly, out of frame, bad composition, low resolution, grainy, pixelated, noisy, oversaturated, undersaturated, (worst quality, low quality:1.3), (bad hands, missing fingers:1.2)" ) negative_prompt_3 = gr.Text( label="Negative prompt 3", max_lines=1, placeholder="Enter a third negative prompt", visible=True, value="(worst quality, low quality:1.3), (bad anatomy, bad hands, missing fingers, extra digit, fewer digits:1.2), (blurry:1.1), cropped, watermark, text, signature, logo, jpeg artifacts, (ugly, deformed, disfigured:1.2), (poorly drawn:1.2), mutated, extra limbs, (bad proportions, gross proportions:1.2), (malformed limbs, missing arms, missing legs, extra arms, extra legs:1.2), (fused fingers, too many fingers, long neck:1.2), (unnatural body, unnatural pose:1.1), out of frame, (bad composition, poorly composed:1.1), (oversaturated, undersaturated:1.1), (grainy, pixelated:1.1), (low resolution, noisy:1.1), (unrealistic, distorted:1.1), (extra fingers, mutated hands, poorly drawn hands, bad hands:1.3), (missing fingers:1.3)" ) num_iterations = gr.Number( value=1000, label="Number of Iterations") with gr.Row(): width = gr.Slider( label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=768, ) height = gr.Slider( label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=768, ) guidance_scale = gr.Slider( label="Guidance scale", minimum=0.0, maximum=30.0, step=0.1, value=4.2, ) num_inference_steps = gr.Slider( label="Number of inference steps", minimum=1, maximum=500, step=1, value=50, ) gr.on( triggers=[run_button_30.click, prompt.submit], fn=infer_30, inputs=[ prompt, negative_prompt_1, negative_prompt_2, negative_prompt_3, width, height, guidance_scale, num_inference_steps, ], outputs=[result, expanded_prompt_output], ) gr.on( triggers=[run_button_60.click, prompt.submit], fn=infer_60, inputs=[ prompt, negative_prompt_1, negative_prompt_2, negative_prompt_3, width, height, guidance_scale, num_inference_steps, ], outputs=[result, expanded_prompt_output], ) gr.on( triggers=[run_button_90.click, prompt.submit], fn=infer_90, inputs=[ prompt, negative_prompt_1, negative_prompt_2, negative_prompt_3, width, height, guidance_scale, num_inference_steps, ], outputs=[result, expanded_prompt_output], ) gr.on( triggers=[run_button_100.click, prompt.submit], fn=infer_100, inputs=[ prompt, negative_prompt_1, negative_prompt_2, negative_prompt_3, width, height, guidance_scale, num_inference_steps, ], outputs=[result, expanded_prompt_output], ) if __name__ == "__main__": demo.launch()