Create deepsite_backend.py

deepsite_backend.py

@@ -0,0 +1,122 @@
+import gradio as gr
+import torch
+from PIL import Image
+from transformers import BlipProcessor, BlipForConditionalGeneration, CLIPProcessor, CLIPModel
+from torchvision import transforms
+
+# Geräteeinstellung
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+# BLIP-Modelle
+blip_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
+blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base").to(device)
+
+# CLIP-Modelle
+clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to(device)
+clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+# Platzhalter für DeepDanbooru
+
+def danbooru_tagging(image):
+    return "1girl, bodysuit, sitting, wooden floor, solo"
+
+def generate_blip_caption(image):
+    raw_image = Image.open(image).convert("RGB")
+    inputs = blip_processor(raw_image, return_tensors="pt").to(device)
+    out = blip_model.generate(**inputs)
+    caption = blip_processor.decode(out[0], skip_special_tokens=True)
+    return caption
+
+def generate_clip_prompt(image, detail_level):
+    raw_image = Image.open(image).convert("RGB")
+    inputs = clip_processor(images=raw_image, return_tensors="pt").to(device)
+    outputs = clip_model.get_image_features(**inputs)
+    base_prompt = "a woman in a bodysuit on wooden floor"
+    if detail_level >= 4:
+        return base_prompt + ", cinematic lighting, ultra detailed, HDR"
+    elif detail_level == 3:
+        return base_prompt + ", moody atmosphere"
+    elif detail_level == 2:
+        return base_prompt + ", minimal shadows"
+    else:
+        return base_prompt
+
+def get_output(image, output_type, style, detail_level, tags, model_choice):
+    if model_choice == "BLIP":
+        if output_type == "Detailed Description":
+            return generate_blip_caption(image)
+        elif output_type == "Short Caption":
+            return generate_blip_caption(image).split(",")[0]
+        elif output_type == "Model Training Data":
+            return generate_blip_caption(image).lower().replace(" ", "_")
+        elif output_type == "AI Prompt":
+            return generate_clip_prompt(image, detail_level)
+    elif model_choice == "CLIP":
+        return generate_clip_prompt(image, detail_level)
+    elif model_choice == "DeepDanbooru":
+        return danbooru_tagging(image)
+    elif model_choice == "NSFW Detector":
+        return "(Simulierter NSFW-Klassifikator: Ergebnis nicht implementiert)"
+    return "[Keine gültige Auswahl getroffen]"
+
+with gr.Blocks(css="style.css") as app:
+    gr.Markdown("# NSFW Image to Text Generator ✨")
+
+    with gr.Row():
+        with gr.Column():
+            img = gr.Image(type="filepath", label="Upload Image")
+            output_type = gr.Dropdown([
+                "Detailed Description", "Short Caption", "Tag List", "AI Prompt", "Model Training Data"], 
+                label="Output Type", value="Detailed Description")
+            style = gr.Dropdown([
+                "Neutral", "Erotic", "Artistic", "Technical", "Literary", "BDSM", "Fetish"], 
+                label="Style", value="Neutral")
+            detail = gr.Slider(1, 5, step=1, value=3, label="Detail Level")
+            tags = gr.Textbox(label="Custom Tags (comma separated)")
+            model_choice = gr.Radio([
+                "CLIP", "BLIP", "DeepDanbooru", "NSFW Detector"], 
+                label="AI Model", value="BLIP")
+            btn_generate = gr.Button("Generate Text")
+
+        with gr.Column():
+            output = gr.Textbox(label="Generated Output", lines=8)
+            gr.Button("Enhance")
+            gr.Button("Shorten")
+            gr.Button("Rewrite")
+
+    btn_generate.click(get_output, 
+        inputs=[img, output_type, style, detail, tags, model_choice], 
+        outputs=output)
+
+    with gr.Tab("Prompt Tools"):
+        prompt_input = gr.Textbox(label="Prompt Builder")
+        btn_optimize = gr.Button("Optimize Prompt")
+        btn_random = gr.Button("Randomize")
+        optimized_output = gr.Textbox(label="Optimized Prompt")
+        btn_optimize.click(lambda p: p + ", ultra detailed", inputs=prompt_input, outputs=optimized_output)
+        btn_random.click(lambda: "a cyberpunk alley at night", outputs=optimized_output)
+
+    with gr.Tab("Training Data"):
+        btn_tags = gr.Button("Generate Tags")
+        tags_out = gr.Textbox(label="Training Tags")
+        btn_tags.click(lambda: "1girl, solo, black bodysuit, sitting", outputs=tags_out)
+
+        caption_mode = gr.Dropdown([
+            "Basic Caption", "Detailed Description", "Booru Style", "Natural Language"], 
+            label="Caption Generation")
+        btn_caption = gr.Button("Generate Caption")
+        caption_out = gr.Textbox(label="Training Caption")
+        btn_caption.click(lambda mode: {
+            "Basic Caption": "A woman posing for a photo",
+            "Detailed Description": "A woman in a futuristic city wearing a sleek bodysuit.",
+            "Booru Style": "1girl, bodysuit, city, night",
+            "Natural Language": "She stands still beneath neon lights, calm yet focused."
+        }.get(mode, ""), inputs=caption_mode, outputs=caption_out)
+
+        trigger_word = gr.Textbox(label="Trigger Word")
+        trigger_class = gr.Textbox(label="Class")
+        btn_lora = gr.Button("Prepare LoRA Training Data")
+        lora_out = gr.Textbox(label="LoRA Output")
+        btn_lora.click(lambda t, c: f"LoRA: {t}, class: {c}", inputs=[trigger_word, trigger_class], outputs=lora_out)
+
+app.launch()