Update app.py

app.py

@@ -7,8 +7,8 @@ from PIL import Image
 # === AUTO HARDWARE ===
 device = "cuda" if torch.cuda.is_available() else "cpu"
 torch_dtype = torch.float16 if device == "cuda" else torch.float32
-IMG_SIZE = 256  # FEST AUF 256 – auch auf CPU (kein Pixelmüll!)
-STEPS = 50 if device == "cuda" else 35
+IMG_SIZE = 512 if device == "cuda" else 384
+STEPS = 50 if device == "cuda" else 40
 
 # === GLOBALE VARIABLEN ===
 pipe_txt2img = None
@@ -19,10 +19,12 @@ def load_txt2img():
     if pipe_txt2img is None:
         print("Lade Text-to-Image Pipeline...")
         pipe_txt2img = StableDiffusionPipeline.from_pretrained(
-            "runwayml/stable-diffusion-v1-5", torch_dtype=torch_dtype
+            "runwayml/stable-diffusion-v1-5", 
+            torch_dtype=torch_dtype,
+            # Kompatibilität mit diffusers 0.29.2
+            use_safetensors=True
         ).to(device)
-        if device == "cpu":
-            pipe_txt2img.enable_attention_slicing()
+        pipe_txt2img.enable_attention_slicing()
     return pipe_txt2img
 
 def load_img2img():
@@ -30,46 +32,71 @@ def load_img2img():
     if pipe_img2img is None:
         print("Lade Img2Img Pipeline...")
         pipe_img2img = StableDiffusionImg2ImgPipeline.from_pretrained(
-            "runwayml/stable-diffusion-v1-5", torch_dtype=torch_dtype
+            "runwayml/stable-diffusion-v1-5", 
+            torch_dtype=torch_dtype,
+            use_safetensors=True
         ).to(device)
-        if device == "cpu":
-            pipe_img2img.enable_attention_slicing()
+        pipe_img2img.enable_attention_slicing()
     return pipe_img2img
 
 # === FUNKTIONEN ===
 def text_to_image(prompt):
-    pipe = load_txt2img()
-    img = pipe(prompt, height=IMG_SIZE, width=IMG_SIZE, num_inference_steps=STEPS).images[0]
-    return img.resize((512, 512), Image.BICUBIC)  # Perfekt scharf!
+    try:
+        pipe = load_txt2img()
+        result = pipe(
+            prompt=prompt,
+            height=IMG_SIZE,
+            width=IMG_SIZE,
+            num_inference_steps=STEPS,
+            guidance_scale=7.5,
+        )
+        return result.images[0]
+    except Exception as e:
+        print(f"Fehler in text_to_image: {e}")
+        return None
 
 def img_to_image(image, prompt="", neg_prompt="", strength=0.75):
-    pipe = load_img2img()
-    img = image.convert("RGB").resize((IMG_SIZE, IMG_SIZE))
-    result = pipe(
-        prompt=prompt or "snowy winter landscape",
-        negative_prompt=neg_prompt or "blur furniture, grass",
-        image=img,
-        strength=strength,
-        guidance_scale=8.5,
-        num_inference_steps=STEPS
-    ).images[0]
-    return result.resize((512, 512), Image.BICUBIC)  # Perfekt scharf!
+    try:
+        pipe = load_img2img()
+        img = image.convert("RGB").resize((IMG_SIZE, IMG_SIZE))
+        result = pipe(
+            prompt=prompt or "snowy winter landscape",
+            negative_prompt=neg_prompt or "blur, furniture, grass, table",
+            image=img,
+            strength=strength,
+            guidance_scale=7.5,
+            num_inference_steps=STEPS
+        )
+        return result.images[0]
+    except Exception as e:
+        print(f"Fehler in img_to_image: {e}")
+        return None
 
 # === UI ===
 with gr.Blocks() as app:
-    gr.Markdown("# Text-to-Image + Img2Img\n**Perfekt scharf | 256→512 | Unter 50 GB**")
+    gr.Markdown("# Text-to-Image + Img2Img\n**Optimierte Version für bessere Ergebnisse**")
     
     with gr.Tab("Text to Image"):
-        txt_in = gr.Textbox(placeholder="z.B. 'Ein Schneemann im Wald'", lines=2)
-        txt_out = gr.Image(label="Ergebnis (512x512)")
-        gr.Button("Generieren").click(text_to_image, txt_in, txt_out)
+        gr.Markdown("### Tipp: Verwende englische Prompts für beste Ergebnisse")
+        txt_in = gr.Textbox(
+            placeholder="z.B. 'a snowman in a forest, winter, snowy trees, highly detailed'",
+            lines=2,
+            value="a snowman in a forest, winter, snowy trees, highly detailed"
+        )
+        txt_out = gr.Image(label="Ergebnis")
+        txt_btn = gr.Button("Generieren")
+        txt_btn.click(text_to_image, txt_in, txt_out)
     
     with gr.Tab("Bild zu Schneelandschaft"):
         img_in = gr.Image(type="pil")
-        prompt = gr.Textbox(value="girl hugging snowman in snowy landscape, realistic", lines=2)
-        neg = gr.Textbox(value="table, tarp, grass, blur", lines=2)
+        prompt = gr.Textbox(
+            value="girl hugging snowman in snowy landscape, realistic, winter",
+            lines=2
+        )
+        neg = gr.Textbox(value="table, tarp, grass, blur, furniture", lines=2)
         strength = gr.Slider(0.6, 0.9, 0.75)
-        img_out = gr.Image(label="Ergebnis (512x512)")
-        gr.Button("Generieren").click(img_to_image, [img_in, prompt, neg, strength], img_out)
+        img_out = gr.Image(label="Ergebnis")
+        img_btn = gr.Button("Generieren")
+        img_btn.click(img_to_image, [img_in, prompt, neg, strength], img_out)
 
 app.launch(share=True)