app.py

import gradio as gr
import torch
from transformers import CLIPProcessor, CLIPModel
from torch import nn
import numpy as np
import PIL
from PIL import Image
from torchvision import transforms

# Load CLIP model and processor
model_name = "openai/clip-vit-base-patch16"
clip_model = CLIPModel.from_pretrained(model_name)
clip_processor = CLIPProcessor.from_pretrained(model_name)

# Generate a random noise tensor (this will be transformed into an image)
def generate_image_from_text(text_input):
    # Preprocess the input text for CLIP model
    inputs = clip_processor(text=text_input, return_tensors="pt", padding=True)

    # Extract image-text features using CLIP
    text_features = clip_model.get_text_features(**inputs)

    # Create a simple GAN-like generator using a random noise tensor
    class SimpleGenerator(nn.Module):
        def __init__(self):
            super(SimpleGenerator, self).__init__()
            self.fc = nn.Linear(512, 256*256*3)  # Adjust output size to match image dimensions
            self.relu = nn.ReLU()

        def forward(self, z):
            x = self.fc(z)
            x = self.relu(x)
            x = x.view(-1, 3, 256, 256)  # Reshape to match image shape
            return x

    # Initialize the generator
    generator = SimpleGenerator()

    # Generate random noise based on the text features
    random_input = torch.randn(1, 512)  # Matching CLIP output size (text_features shape)
    generated_image_tensor = generator(random_input)
    
    # Convert generated image tensor to PIL Image
    generated_image = generated_image_tensor.squeeze().permute(1, 2, 0).detach().numpy()
    generated_image = np.clip(generated_image, 0, 1)  # Normalize pixel values
    generated_image = (generated_image * 255).astype(np.uint8)
    generated_image = Image.fromarray(generated_image)

    return generated_image

# Gradio interface
iface = gr.Interface(fn=generate_image_from_text, inputs="text", outputs="image", live=True)

iface.launch()