sdascdv

app.py

@@ -1,14 +1,14 @@
 import os
 import torch
 import gradio as gr
-from diffusers import StableDiffusionImg2ImgPipeline, StableDiffusionPipeline
+from diffusers import StableDiffusionImg2ImgPipeline
 from PIL import Image
 
 # --- Configuration ---
 SPACE_TITLE = "🎨 Enhanced Studio Ghibli AI Art Generator"
 SPACE_DESCRIPTION = "Upload a portrait or a photo and transform it into a breathtaking Studio Ghibli-style masterpiece! Improved model and prompting for better results."
 MAIN_MODEL_ID = "nitrosocke/Ghibli-Diffusion"
-STYLE_MODEL_ID = "sayakpaul/sd-anime-diff"  # Experiment with other anime style models
+# STYLE_MODEL_ID = "sayakpaul/sd-anime-diff"  # Removed the problematic style model
 STRENGTH = 0.60  # Adjust for better balance between input and style
 GUIDANCE_SCALE = 7.5  # Increased for better prompt adherence
 NUM_INFERENCE_STEPS = 30  # Increased for potentially higher quality
@@ -42,25 +42,14 @@ img2img_pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
     low_cpu_mem_usage=True
 ).to(device)
 
-# Load a separate Stable Diffusion model for generating initial style (optional, but can help)
-# You can comment this out if you only want to rely on the img2img model
-style_pipe = StableDiffusionPipeline.from_pretrained(
-    STYLE_MODEL_ID,
-    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
-    use_safetensors=True,
-    low_cpu_mem_usage=True
-).to(device)
-
 # --- Optimization (Conditional for CUDA) ---
 if device == "cuda":
     try:
         img2img_pipe.enable_xformers_memory_efficient_attention()
-        style_pipe.enable_xformers_memory_efficient_attention()
         print("✅ xFormers enabled!")
     except Exception as e:
         print(f"⚠️ xFormers not available: {e}")
     img2img_pipe.enable_model_cpu_offload()
-    style_pipe.enable_model_cpu_offload()
 pipe.enable_vae_slicing()
 pipe.enable_attention_slicing()
 
@@ -72,11 +61,6 @@ def transform_image(input_image):
     try:
         input_image = input_image.resize(INPUT_IMAGE_SIZE)
 
-        # Generate an initial stylized image using the style model (optional)
-        # initial_style_prompt = f"{PROMPT_PREFIX} portrait of a person"
-        # initial_style_image = style_pipe(prompt=initial_style_prompt, negative_prompt=NEGATIVE_PROMPT, num_inference_steps=NUM_INFERENCE_STEPS // 2).images[0]
-
-        # Use the input image directly with the img2img pipeline
         output = img2img_pipe(
             prompt=PROMPT_PREFIX + "portrait of a person", # Adjust prompt based on input
             image=input_image,