app.py

#!/usr/bin/env python
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
import spaces

import os
import random
import uuid
import gradio as gr
import numpy as np
from PIL import Image
import torch
#import diffusers
from diffusers import AutoencoderKL, StableDiffusionXLPipeline, UNet2DConditionModel
from diffusers import EulerAncestralDiscreteScheduler
from typing import Tuple
import paramiko
import datetime
#from diffusers import DPMSolverSDEScheduler
from diffusers.models.attention_processor import AttnProcessor2_0
from transformers import CLIPTextModelWithProjection, CLIPTextModel
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"
torch.set_float32_matmul_precision("highest")

FTP_HOST = "1ink.us"
FTP_USER = "ford442"
FTP_PASS = os.getenv("FTP_PASS")
FTP_DIR = "1ink.us/stable_diff/"  # Remote directory on FTP server

DESCRIPTIONXX = """
    ## ⚡⚡⚡⚡ REALVISXL V5.0 BF16 (Tester F) ⚡⚡⚡⚡
"""

examples = [
    "Many apples splashed with drops of water within a fancy bowl 4k, hdr  --v 6.0 --style raw",
    "A profile photo of a dog, brown background, shot on Leica M6 --ar 128:85 --v 6.0 --style raw",
]

MODEL_OPTIONS = {
    "REALVISXL V5.0 BF16": "ford442/RealVisXL_V5.0_BF16",
}

MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "4096"))
BATCH_SIZE = int(os.getenv("BATCH_SIZE", "1"))

style_list = [
    {
        "name": "3840 x 2160",
        "prompt": "hyper-realistic 8K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
        "negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly",
    },
    {
        "name": "2560 x 1440",
        "prompt": "hyper-realistic 4K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
        "negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly",
    },
    {
        "name": "HD+",
        "prompt": "hyper-realistic 2K image of {prompt}. ultra-detailed, lifelike, high-resolution, sharp, vibrant colors, photorealistic",
        "negative_prompt": "cartoonish, low resolution, blurry, simplistic, abstract, deformed, ugly",
    },
    {
        "name": "Style Zero",
        "prompt": "{prompt}",
        "negative_prompt": "",
    },
]

styles = {k["name"]: (k["prompt"], k["negative_prompt"]) for k in style_list}
DEFAULT_STYLE_NAME = "Style Zero"
STYLE_NAMES = list(styles.keys())
HF_TOKEN = os.getenv("HF_TOKEN")
os.putenv("HF_HUB_ENABLE_HF_TRANSFER","1")
os.environ["SAFETENSORS_FAST_GPU"] = "1"

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    
def load_and_prepare_model():
    unet = UNet2DConditionModel.from_pretrained("ford442/RealVisXL_V5.0_BF16", low_cpu_mem_usage=False, subfolder='unet', upcast_attention=True, attention_type='gated-text-image', token=True)
    #vaeRV = AutoencoderKL.from_pretrained("SG161222/RealVisXL_V5.0", subfolder='vae', safety_checker=None, use_safetensors=True, token=True)
    #vaeXL = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae", safety_checker=None, use_safetensors=False, low_cpu_mem_usage=False, torch_dtype=torch.float32, token=True) #.to(device).to(torch.bfloat16) #.to(device=device, dtype=torch.bfloat16)
    vaeXL = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae", low_cpu_mem_usage=False, safety_checker=None, use_safetensors=False, torch_dtype=torch.float32, token=True) #.to(device).to(torch.bfloat16) #.to(device=device, dtype=torch.bfloat16)
    #sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler',beta_schedule="scaled_linear", beta_start=0.00085, beta_end=0.012, steps_offset=1,use_karras_sigmas=True)
    #sched = DPMSolverSDEScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler')
    #sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler',beta_schedule="scaled_linear", token=True) #, beta_start=0.00085, beta_end=0.012, steps_offset=1,use_karras_sigmas=True, token=True)
    #sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler',beta_schedule="scaled_linear")
    pipe = StableDiffusionXLPipeline.from_pretrained(
        'ford442/RealVisXL_V5.0_BF16',
        #torch_dtype=torch.bfloat16,
        token=True,
        add_watermarker=False,
        #text_encoder=None,
        #text_encoder_2=None,
        unet=unet,
        vae=None,
    )
    #pipe.vae = vaeXL #.to(torch.bfloat16)
    #pipe.scheduler = sched
    #pipe.vae.do_resize=False
    #pipe.vae.vae_scale_factor=8
    #pipe.to(device)
    #pipe.to(torch.bfloat16)
    pipe.unet.fuse_qkv_projections()
    print(f'init noise scale: {pipe.scheduler.init_noise_sigma}')
    pipe.watermark=None
    pipe.safety_checker=None
    #pipe.unet.to(memory_format=torch.channels_last)
    #pipe.enable_vae_tiling()
    pipe.to(device=device, dtype=torch.bfloat16)
    pipe.vae = vaeXL.to(device) #.to('cpu') #.to(torch.bfloat16)
    
    pipe.unet.set_attn_processor(AttnProcessor2_0())
    pipe.vae.set_default_attn_processor()
    return pipe
    
pipe = load_and_prepare_model()

text_encoder=CLIPTextModel.from_pretrained('ford442/RealVisXL_V5.0_BF16', low_cpu_mem_usage=False, subfolder='text_encoder',token=True)#.to(device=device, dtype=torch.bfloat16)
text_encoder_2=CLIPTextModelWithProjection.from_pretrained('ford442/RealVisXL_V5.0_BF16', low_cpu_mem_usage=False, subfolder='text_encoder_2',token=True)#.to(device=device, dtype=torch.bfloat16)

MAX_SEED = np.iinfo(np.int32).max

neg_prompt_2 = " 'non-photorealistic':1.5, 'unrealistic skin','unattractive face':1.3, 'low quality':1.1, ('dull color scheme', 'dull colors', 'digital noise':1.2),'amateurish', 'poorly drawn face':1.3, 'poorly drawn', 'distorted face', 'low resolution', 'simplistic' "

def upload_to_ftp(filename):
    try:
        transport = paramiko.Transport((FTP_HOST, 22))
        if filename.endswith(".txt"):
            destination_path=FTP_DIR+'/txt/'+filename
        else:
            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}")

def apply_style(style_name: str, positive: str, negative: str = "") -> Tuple[str, str]:
    if style_name in styles:
        p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME])
    else:
        p, n = styles[DEFAULT_STYLE_NAME]
    if not negative:
        negative = ""
    return p.replace("{prompt}", positive), n + negative
    
def save_image(img):
    unique_name = str(uuid.uuid4()) + ".png"
    img.save(unique_name,optimize=False,compress_level=0)
    return unique_name

def uploadNote(prompt,num_inference_steps,guidance_scale,timestamp):
    filename= f'tst_F_{timestamp}.txt'
    with open(filename, "w") as f:
        f.write(f"Realvis 5.0 (Tester F) \n")
        f.write(f"Date/time: {timestamp} \n")
        f.write(f"Prompt: {prompt} \n")
        f.write(f"Steps: {num_inference_steps} \n")
        f.write(f"Guidance Scale: {guidance_scale} \n")
        f.write(f"SPACE SETUP: \n")
        f.write(f"fuse_qkv_projections  \n")
    upload_to_ftp(filename) 
    
@spaces.GPU(duration=30)
def generate_30(
    prompt: str,
    prompt2: str,
    negative_prompt: str = "",
    use_negative_prompt: bool = False,
    style_selection: str = "",
    width: int = 768,
    height: int = 768,
    guidance_scale: float = 4,
    num_inference_steps: int = 125,
    txt_strength: float = 1.0,
    use_resolution_binning: bool = True, 
    progress=gr.Progress(track_tqdm=True)  # Add progress as a keyword argument
):
    seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device='cuda').manual_seed(seed)
    pipe.text_encoder=text_encoder.to(device=device, dtype=torch.bfloat16)
    pipe.text_encoder_2=text_encoder_2.to(device=device, dtype=torch.bfloat16)

    pooled_prompt_embeds_list=[]
    prompt_embeds_list=[]
    
    text_inputs1 = pipe.tokenizer(
                    prompt,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )
    
    text_input_ids1 = text_inputs1.input_ids
    
    text_inputs2 = pipe.tokenizer(
                    prompt2,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )
    
    text_input_ids2 = text_inputs2.input_ids

    text_inputs1b = pipe.tokenizer_2(
                    prompt,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )

    text_input_ids1b = text_inputs1b.input_ids

    text_inputs2b = pipe.tokenizer_2(
                    prompt2,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )
    text_input_ids2b = text_inputs2b.input_ids

    # 2. Encode with the two text encoders
    prompt_embeds_a = pipe.text_encoder(text_input_ids1.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_a = prompt_embeds_a[0][:, -1, :]  # Pooled output from encoder 1
    print('pooled shape 1: ', pooled_prompt_embeds_a.shape)
    prompt_embeds_a = prompt_embeds_a.hidden_states[-2]  # Penultimate hidden state from encoder 1
    print('encoder shape: ', prompt_embeds_a.shape)
    prompt_embeds_b = pipe.text_encoder(text_input_ids2.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_b = prompt_embeds_b[0][:, -1, :]  # Pooled output from encoder 1
    prompt_embeds_b = prompt_embeds_b.hidden_states[-2]  # Penultimate hidden state from encoder 1

    prompt_embeds_a2 = pipe.text_encoder_2(text_input_ids1b.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_a2 = prompt_embeds_a2[0]  # Pooled output from encoder 2
    print('pooled shape 2: ', pooled_prompt_embeds_a2.shape)
    prompt_embeds_a2 = prompt_embeds_a2.hidden_states[-2]  # Penultimate hidden state from encoder 2
    print('encoder shape2: ', prompt_embeds_a2.shape)
    prompt_embeds_b2 = pipe.text_encoder_2(text_input_ids2b.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_b2 = prompt_embeds_b2[0]  # Pooled output from encoder 2
    prompt_embeds_b2 = prompt_embeds_b2.hidden_states[-2]  # Penultimate hidden state from encoder 2

    # 3. Concatenate the embeddings 
    prompt_embeds = torch.cat([prompt_embeds_a, prompt_embeds_b])
    print('catted shape: ', prompt_embeds.shape)
    pooled_prompt_embeds = torch.cat([pooled_prompt_embeds_a, pooled_prompt_embeds_b])
    print('catted pooled shape: ', pooled_prompt_embeds.shape)
    pooled_prompt_embeds = torch.mean(pooled_prompt_embeds,dim=0,keepdim=True)
    print('meaned pooled shape: ', pooled_prompt_embeds.shape)

    # 4. (Optional) Average the pooled embeddings
    prompt_embeds = torch.mean(prompt_embeds,dim=0,keepdim=True)
    print('averaged shape: ', prompt_embeds.shape)

    # 3. Concatenate the text_encoder_2 embeddings 
    prompt_embeds2 = torch.cat([prompt_embeds_a2, prompt_embeds_b2])
    print('catted shape2: ', prompt_embeds2.shape)
    prompt_embeds2 = torch.mean(prompt_embeds2,dim=0,keepdim=True)
    print('averaged shape2: ', prompt_embeds.shape)

    pooled_prompt_embeds2 = torch.cat([pooled_prompt_embeds_a2, pooled_prompt_embeds_b2])
    print('catted pooled shape 2: ', pooled_prompt_embeds2.shape)
    pooled_prompt_embeds2 = torch.mean(pooled_prompt_embeds2,dim=0,keepdim=True)
    print('pooled meaned shape 2: ', pooled_prompt_embeds2.shape)
    pooled_prompt_embeds = pooled_prompt_embeds2 #torch.cat([pooled_prompt_embeds, pooled_prompt_embeds2],dim=1)
    print('catted combined meaned pooled shape: ', pooled_prompt_embeds.shape)
  
    prompt_embeds = torch.cat([prompt_embeds, prompt_embeds2],dim=-1)
    print('catted combined meaned shape: ', prompt_embeds.shape)

    prompt_embeds = prompt_embeds * txt_strength
    pooled_prompt_embeds = pooled_prompt_embeds * txt_strength
    
    options = {
        #"prompt": prompt,
        "prompt_embeds": prompt_embeds,
        "pooled_prompt_embeds": pooled_prompt_embeds,
        "negative_prompt": negative_prompt,
        "negative_prompt_2": neg_prompt_2,
        "width": width,
        "height": height,
        "guidance_scale": guidance_scale,
        "num_inference_steps": num_inference_steps,
        "generator": generator,
        "output_type": "pil",
    }
    if use_resolution_binning:
        options["use_resolution_binning"] = True
    images = []
    timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
    uploadNote(prompt,num_inference_steps,guidance_scale,timestamp)
    batch_options = options.copy()
    rv_image = pipe(**batch_options).images[0]
    sd_image_path = f"rv50_F_{timestamp}.png"
    rv_image.save(sd_image_path,optimize=False,compress_level=0)
    upload_to_ftp(sd_image_path)    
    unique_name = str(uuid.uuid4()) + ".png"  
    os.symlink(sd_image_path, unique_name)  
    return [unique_name]

@spaces.GPU(duration=60)
def generate_60(
    prompt: str,
    prompt2: str,
    negative_prompt: str = "",
    use_negative_prompt: bool = False,
    style_selection: str = "",
    width: int = 768,
    height: int = 768,
    guidance_scale: float = 4,
    num_inference_steps: int = 125,
    txt_strength: float = 1.0,
    use_resolution_binning: bool = True, 
    progress=gr.Progress(track_tqdm=True)  # Add progress as a keyword argument
):
    seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device='cuda').manual_seed(seed)
    pipe.text_encoder=text_encoder.to(device=device, dtype=torch.bfloat16)
    pipe.text_encoder_2=text_encoder_2.to(device=device, dtype=torch.bfloat16)

    pooled_prompt_embeds_list=[]
    prompt_embeds_list=[]
    
    text_inputs1 = pipe.tokenizer(
                    prompt,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )
    
    text_input_ids1 = text_inputs1.input_ids
    
    text_inputs2 = pipe.tokenizer(
                    prompt2,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )
    
    text_input_ids2 = text_inputs2.input_ids

    text_inputs1b = pipe.tokenizer_2(
                    prompt,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )

    text_input_ids1b = text_inputs1b.input_ids

    text_inputs2b = pipe.tokenizer_2(
                    prompt2,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )
    text_input_ids2b = text_inputs2b.input_ids

    # 2. Encode with the two text encoders
    prompt_embeds_a = pipe.text_encoder(text_input_ids1.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_a = prompt_embeds_a[0][:, -1, :]  # Pooled output from encoder 1
    print('pooled shape 1: ', pooled_prompt_embeds_a.shape)
    prompt_embeds_a = prompt_embeds_a.hidden_states[-2]  # Penultimate hidden state from encoder 1
    print('encoder shape: ', prompt_embeds_a.shape)
    prompt_embeds_b = pipe.text_encoder(text_input_ids2.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_b = prompt_embeds_b[0][:, -1, :]  # Pooled output from encoder 1
    prompt_embeds_b = prompt_embeds_b.hidden_states[-2]  # Penultimate hidden state from encoder 1

    prompt_embeds_a2 = pipe.text_encoder_2(text_input_ids1b.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_a2 = prompt_embeds_a2[0]  # Pooled output from encoder 2
    print('pooled shape 2: ', pooled_prompt_embeds_a2.shape)
    prompt_embeds_a2 = prompt_embeds_a2.hidden_states[-2]  # Penultimate hidden state from encoder 2
    print('encoder shape2: ', prompt_embeds_a2.shape)
    prompt_embeds_b2 = pipe.text_encoder_2(text_input_ids2b.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_b2 = prompt_embeds_b2[0]  # Pooled output from encoder 2
    prompt_embeds_b2 = prompt_embeds_b2.hidden_states[-2]  # Penultimate hidden state from encoder 2

    # 3. Concatenate the embeddings 
    prompt_embeds = torch.cat([prompt_embeds_a, prompt_embeds_b])
    print('catted shape: ', prompt_embeds.shape)
    pooled_prompt_embeds = torch.cat([pooled_prompt_embeds_a, pooled_prompt_embeds_b])
    print('catted pooled shape: ', pooled_prompt_embeds.shape)
    pooled_prompt_embeds = torch.mean(pooled_prompt_embeds,dim=0,keepdim=True)
    print('meaned pooled shape: ', pooled_prompt_embeds.shape)

    # 4. (Optional) Average the pooled embeddings
    prompt_embeds = torch.mean(prompt_embeds,dim=0,keepdim=True)
    print('averaged shape: ', prompt_embeds.shape)

    # 3. Concatenate the text_encoder_2 embeddings 
    prompt_embeds2 = torch.cat([prompt_embeds_a2, prompt_embeds_b2])
    print('catted shape2: ', prompt_embeds2.shape)
    prompt_embeds2 = torch.mean(prompt_embeds2,dim=0,keepdim=True)
    print('averaged shape2: ', prompt_embeds.shape)

    pooled_prompt_embeds2 = torch.cat([pooled_prompt_embeds_a2, pooled_prompt_embeds_b2])
    print('catted pooled shape 2: ', pooled_prompt_embeds2.shape)
    pooled_prompt_embeds2 = torch.mean(pooled_prompt_embeds2,dim=0,keepdim=True)
    print('pooled meaned shape 2: ', pooled_prompt_embeds2.shape)
    pooled_prompt_embeds = pooled_prompt_embeds2 #torch.cat([pooled_prompt_embeds, pooled_prompt_embeds2],dim=1)
    print('catted combined meaned pooled shape: ', pooled_prompt_embeds.shape)
  
    prompt_embeds = torch.cat([prompt_embeds, prompt_embeds2],dim=-1)
    print('catted combined meaned shape: ', prompt_embeds.shape)

    prompt_embeds = prompt_embeds * txt_strength
    pooled_prompt_embeds = pooled_prompt_embeds * txt_strength
    
    options = {
        #"prompt": prompt,
        "prompt_embeds": prompt_embeds,
        "pooled_prompt_embeds": pooled_prompt_embeds,
        "negative_prompt": negative_prompt,
        "negative_prompt_2": neg_prompt_2,
        "width": width,
        "height": height,
        "guidance_scale": guidance_scale,
        "num_inference_steps": num_inference_steps,
        "generator": generator,
        "output_type": "pil",
    }
    if use_resolution_binning:
        options["use_resolution_binning"] = True
    images = []
    timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
    uploadNote(prompt,num_inference_steps,guidance_scale,timestamp)
    batch_options = options.copy()
    rv_image = pipe(**batch_options).images[0]
    sd_image_path = f"rv50_F_{timestamp}.png"
    rv_image.save(sd_image_path,optimize=False,compress_level=0)
    upload_to_ftp(sd_image_path)    
    unique_name = str(uuid.uuid4()) + ".png"  
    os.symlink(sd_image_path, unique_name)  
    return [unique_name]
    
@spaces.GPU(duration=90)
def generate_90(
    prompt: str,
    prompt2: str,
    negative_prompt: str = "",
    use_negative_prompt: bool = False,
    style_selection: str = "",
    width: int = 768,
    height: int = 768,
    guidance_scale: float = 4,
    num_inference_steps: int = 125,
    txt_strength: float = 1.0,
    use_resolution_binning: bool = True, 
    progress=gr.Progress(track_tqdm=True)  # Add progress as a keyword argument
):
    seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device='cuda').manual_seed(seed)
    pipe.text_encoder=text_encoder.to(device=device, dtype=torch.bfloat16)
    pipe.text_encoder_2=text_encoder_2.to(device=device, dtype=torch.bfloat16)

    pooled_prompt_embeds_list=[]
    prompt_embeds_list=[]
    
    text_inputs1 = pipe.tokenizer(
                    prompt,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )
    
    text_input_ids1 = text_inputs1.input_ids
    
    text_inputs2 = pipe.tokenizer(
                    prompt2,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )
    
    text_input_ids2 = text_inputs2.input_ids

    text_inputs1b = pipe.tokenizer_2(
                    prompt,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )

    text_input_ids1b = text_inputs1b.input_ids

    text_inputs2b = pipe.tokenizer_2(
                    prompt2,
                    padding="max_length",
                    max_length=77,
                    truncation=True,
                    return_tensors="pt",
    )
    text_input_ids2b = text_inputs2b.input_ids

    # 2. Encode with the two text encoders
    prompt_embeds_a = pipe.text_encoder(text_input_ids1.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_a = prompt_embeds_a[0][:, -1, :]  # Pooled output from encoder 1
    print('pooled shape 1: ', pooled_prompt_embeds_a.shape)
    prompt_embeds_a = prompt_embeds_a.hidden_states[-2]  # Penultimate hidden state from encoder 1
    print('encoder shape: ', prompt_embeds_a.shape)
    prompt_embeds_b = pipe.text_encoder(text_input_ids2.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_b = prompt_embeds_b[0][:, -1, :]  # Pooled output from encoder 1
    prompt_embeds_b = prompt_embeds_b.hidden_states[-2]  # Penultimate hidden state from encoder 1

    prompt_embeds_a2 = pipe.text_encoder_2(text_input_ids1b.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_a2 = prompt_embeds_a2[0]  # Pooled output from encoder 2
    print('pooled shape 2: ', pooled_prompt_embeds_a2.shape)
    prompt_embeds_a2 = prompt_embeds_a2.hidden_states[-2]  # Penultimate hidden state from encoder 2
    print('encoder shape2: ', prompt_embeds_a2.shape)
    prompt_embeds_b2 = pipe.text_encoder_2(text_input_ids2b.to(torch.device('cuda')), output_hidden_states=True)
    pooled_prompt_embeds_b2 = prompt_embeds_b2[0]  # Pooled output from encoder 2
    prompt_embeds_b2 = prompt_embeds_b2.hidden_states[-2]  # Penultimate hidden state from encoder 2

    # 3. Concatenate the embeddings 
    prompt_embeds = torch.cat([prompt_embeds_a, prompt_embeds_b])
    print('catted shape: ', prompt_embeds.shape)
    pooled_prompt_embeds = torch.cat([pooled_prompt_embeds_a, pooled_prompt_embeds_b])
    print('catted pooled shape: ', pooled_prompt_embeds.shape)
    pooled_prompt_embeds = torch.mean(pooled_prompt_embeds,dim=0,keepdim=True)
    print('meaned pooled shape: ', pooled_prompt_embeds.shape)

    # 4. (Optional) Average the pooled embeddings
    prompt_embeds = torch.mean(prompt_embeds,dim=0,keepdim=True)
    print('averaged shape: ', prompt_embeds.shape)

    # 3. Concatenate the text_encoder_2 embeddings 
    prompt_embeds2 = torch.cat([prompt_embeds_a2, prompt_embeds_b2])
    print('catted shape2: ', prompt_embeds2.shape)
    prompt_embeds2 = torch.mean(prompt_embeds2,dim=0,keepdim=True)
    print('averaged shape2: ', prompt_embeds.shape)

    pooled_prompt_embeds2 = torch.cat([pooled_prompt_embeds_a2, pooled_prompt_embeds_b2])
    print('catted pooled shape 2: ', pooled_prompt_embeds2.shape)
    pooled_prompt_embeds2 = torch.mean(pooled_prompt_embeds2,dim=0,keepdim=True)
    print('pooled meaned shape 2: ', pooled_prompt_embeds2.shape)
    pooled_prompt_embeds = pooled_prompt_embeds2 #torch.cat([pooled_prompt_embeds, pooled_prompt_embeds2],dim=1)
    print('catted combined meaned pooled shape: ', pooled_prompt_embeds.shape)
  
    prompt_embeds = torch.cat([prompt_embeds, prompt_embeds2],dim=-1)
    print('catted combined meaned shape: ', prompt_embeds.shape)

    prompt_embeds = prompt_embeds * txt_strength
    pooled_prompt_embeds = pooled_prompt_embeds * txt_strength
    
    options = {
        #"prompt": prompt,
        "prompt_embeds": prompt_embeds,
        "pooled_prompt_embeds": pooled_prompt_embeds,
        "negative_prompt": negative_prompt,
        "negative_prompt_2": neg_prompt_2,
        "width": width,
        "height": height,
        "guidance_scale": guidance_scale,
        "num_inference_steps": num_inference_steps,
        "generator": generator,
        "output_type": "pil",
    }
    if use_resolution_binning:
        options["use_resolution_binning"] = True
    images = []
    timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
    uploadNote(prompt,num_inference_steps,guidance_scale,timestamp)
    batch_options = options.copy()
    rv_image = pipe(**batch_options).images[0]
    sd_image_path = f"rv50_F_{timestamp}.png"
    rv_image.save(sd_image_path,optimize=False,compress_level=0)
    upload_to_ftp(sd_image_path)    
    unique_name = str(uuid.uuid4()) + ".png"  
    os.symlink(sd_image_path, unique_name)  
    return [unique_name]

def load_predefined_images1():
    predefined_images1 = [
        "assets/7.png",
        "assets/8.png",
        "assets/9.png",
        "assets/1.png",
        "assets/2.png",
        "assets/3.png",
        "assets/4.png",
        "assets/5.png",
        "assets/6.png",
    ]
    return predefined_images1

css = '''
#col-container {
    margin: 0 auto;
    max-width: 640px;
}
h1{text-align:center}
footer {
    visibility: hidden
}
body {
  background-color: green;
}
'''

with gr.Blocks(theme=gr.themes.Origin(),css=css) as demo:
    gr.Markdown(DESCRIPTIONXX)
    with gr.Row():
        prompt = gr.Text(
            label="Prompt",
            show_label=False,
            max_lines=1,
            placeholder="Enter your prompt",
            container=False,
        )
        prompt2 = gr.Text(
            label="Prompt 2",
            show_label=False,
            max_lines=1,
            placeholder="Enter your prompt",
            container=False,
        )
        run_button_30 = gr.Button("Run 30 Seconds", scale=0)
        run_button_60 = gr.Button("Run 60 Seconds", scale=0)
        run_button_90 = gr.Button("Run 90 Seconds", scale=0)
    result = gr.Gallery(label="Result", columns=1, show_label=False) 

    with gr.Row():

        style_selection = gr.Radio(
            show_label=True,
            container=True,
            interactive=True,
            choices=STYLE_NAMES,
            value=DEFAULT_STYLE_NAME,
            label="Quality Style",
        )
        with gr.Row():
            with gr.Column(scale=1):
                use_negative_prompt = gr.Checkbox(label="Use negative prompt", value=True)
                negative_prompt = gr.Text(
                    label="Negative prompt",
                    max_lines=5,
                    lines=4,
                    placeholder="Enter a negative prompt",
                    value="('deformed', 'distorted', 'disfigured':1.3),'not photorealistic':1.5, 'poorly drawn', 'bad anatomy', 'wrong anatomy', 'extra limb', 'missing limb', 'floating limbs', 'poorly drawn hands', 'poorly drawn feet', 'poorly drawn face':1.3, 'out of frame', 'extra limbs', 'bad anatomy', 'bad art', 'beginner',  'distorted face','amateur'",
                    visible=True,
                )
                text_strength = gr.Slider(
                    label="Text Strength",
                    minimum=0.0,
                    maximum=5.0,
                    step=0.01,
                    value=1.0,
                )
        with gr.Row():
            width = gr.Slider(
                label="Width",
                minimum=448,
                maximum=MAX_IMAGE_SIZE,
                step=64,
                value=768,
            )
            height = gr.Slider(
                label="Height",
                minimum=448,
                maximum=MAX_IMAGE_SIZE,
                step=64,
                value=768,
            )
        with gr.Row():
            guidance_scale = gr.Slider(
                label="Guidance Scale",
                minimum=0.1,
                maximum=30,
                step=0.1,
                value=3.8,
            )
            num_inference_steps = gr.Slider(
                label="Number of inference steps",
                minimum=10,
                maximum=1000,
                step=10,
                value=170,
            )

    gr.Examples(
        examples=examples,
        inputs=prompt,
        cache_examples=False
    )

    use_negative_prompt.change(
        fn=lambda x: gr.update(visible=x),
        inputs=use_negative_prompt,
        outputs=negative_prompt,
        api_name=False,
    )
    
    gr.on(
        triggers=[
            run_button_30.click,
        ],
      #  api_name="generate",  # Add this line
        fn=generate_30,
        inputs=[
            prompt,
            prompt2,
            negative_prompt,
            use_negative_prompt,
            style_selection,
            width,
            height,
            guidance_scale,
            num_inference_steps,
            text_strength,
        ],
        outputs=[result],
    )
    
    gr.on(
        triggers=[
            run_button_60.click,
        ],
      #  api_name="generate",  # Add this line
        fn=generate_60,
        inputs=[
            prompt,
            prompt2,
            negative_prompt,
            use_negative_prompt,
            style_selection,
            width,
            height,
            guidance_scale,
            num_inference_steps,
            text_strength,
        ],
        outputs=[result],
    )
    
    gr.on(
        triggers=[
            run_button_90.click,
        ],
      #  api_name="generate",  # Add this line
        fn=generate_90,
        inputs=[
            prompt,
            prompt2,
            negative_prompt,
            use_negative_prompt,
            style_selection,
            width,
            height,
            guidance_scale,
            num_inference_steps,
            text_strength,
        ],
        outputs=[result],
    )

    gr.Markdown("### REALVISXL V5.0")
    predefined_gallery = gr.Gallery(label="REALVISXL V5.0", columns=3, show_label=False, value=load_predefined_images1())

    gr.Markdown(
    """
    <div style="text-align: justify;">
    ⚡Models used in the playground <a href="https://huggingface.co/SG161222/RealVisXL_V5.0">[REALVISXL V5.0]</a>, <a href="https://huggingface.co/SG161222/RealVisXL_V5.0_Lightning">[REALVISXL V5.0 LIGHTNING]</a> for image generation. Stable Diffusion XL piped (SDXL) model HF. This is the demo space for generating images using the Stable Diffusion XL models, with multiple different variants available.
    </div>
    """)

    gr.Markdown(
    """
    <div style="text-align: justify;">
    ⚡This is the demo space for generating images using Stable Diffusion XL with quality styles, different models, and types. Try the sample prompts to generate higher quality images. Try the sample prompts for generating higher quality images. 
    <a href='https://huggingface.co/spaces/prithivMLmods/Top-Prompt-Collection' target='_blank'>Try prompts</a>.
    </div>
    """)

    gr.Markdown(
    """
    <div style="text-align: justify;">
    ⚠️ Users are accountable for the content they generate and are responsible for ensuring it meets appropriate ethical standards.
    </div>
    """) 

def text_generation(input_text, seed):
    full_prompt = "Text Generator Application by ecarbo"
    return full_prompt
    
title = "Text Generator Demo GPT-Neo"
description = "Text Generator Application by ecarbo"

if __name__ == "__main__":
    demo_interface = demo.queue(max_size=50)  # Remove .launch() here

    text_gen_interface = gr.Interface(
        fn=text_generation,
        inputs=[
            gr.Textbox(lines=1, label="Expand the following prompt to be more detailed and descriptive for image generation: "),
            gr.Number(value=10, label="Enter seed number")
        ],
        outputs=gr.Textbox(label="Text Generated"),
        title=title,
        description=description,
    )

    combined_interface = gr.TabbedInterface([demo_interface, text_gen_interface], ["Image Generation", "Text Generation"])
    combined_interface.launch(show_api=False)