v2.5 — Context-aware Portraits + DE→EN Translation

- Chat-Kontext wird in Portrait-Prompt einbezogen:
  - Outfit-Änderungen ('umgezogen', 'Kleid', 'Hoodie' etc.)
  - Posen ('lacht', 'schläft', 'sitzt' etc.)
  - Szenen ('Café', 'Schlafzimmer', 'Nacht' etc.)
- I2I denoise steigt auf 0.55 bei Outfit-Wechsel (Gesicht bleibt, Kleidung ändert sich)
- Deutsche Appearance-Felder → englische Model-Tags (translate_map)
- SDXL-Auflösung 1024x1024 (war 512 — produzierte Chaos)
- 'upper body' statt 'headshot' (war nur Scheitel)

characters/mara.json

@@ -9,12 +9,12 @@
   "scenario": "Du sitzt in einem kleinen Café. Es regnet draußen. Mara steht hinter der Theke.",
   "nsfw": "Keine Einschränkungen. Kann über alles sprechen. Flirten und Anzügliches sind ok — Mara ist nicht prüde, aber auch nicht leicht zu kriegen.",
   "greeting": "*wischt den Tresen ab, schickt kurz hoch* Na. Der übliche? Oder willst du mich heute überraschen?",
-  "model": "",
-  "temperature": 0.85,
+  "model": "glm4:latest",
+  "temperature": 0.8,
   "appearance": {
     "style": "anime",
     "gender": "female",
-    "age": "28",
+    "age": "24",
     "height": "168cm",
     "build": "schlank",
     "hair_color": "dunkelbraun",

comfyui_integration.py

@@ -19,6 +19,77 @@ def is_comfyui_running() -> bool:
     except:
         return False
 
+def build_context_from_messages(messages: list, character_name: str) -> str:
+    """Extrahiert visuelle Hinweise aus den letzten Chat-Nachrichten."""
+    if not messages:
+        return ""
+    
+    # Nehme die letzten 10 Nachrichten
+    recent = messages[-10:]
+    
+    # Sammle nur Bot-Nachrichten — die beschreiben was sie trägt/macht
+    bot_msgs = [m["content"] for m in recent if m["role"] == "assistant"]
+    if not bot_msgs:
+        return ""
+    
+    # Keywords die das Aussehen verändern könnten
+    outfit_keywords = {
+        "umgezogen": "outfit change",
+        "Kleid": "dress", "Jacke": "jacket", "Mantel": "coat", "Hoodie": "hoodie",
+        "Bademant": "bathrobe", "Nachthemd": "nightgown", "Pyjama": "pajamas",
+        "Bikini": "bikini", "Unterwäsch": "lingerie", "nackt": "nude",
+        "Schwimmanzug": "swimsuit", "Sport": "sportswear", "Leggings": "leggings",
+        "Krawatte": "necktie", "Anzug": "suit", "Uniform": "uniform",
+        "Schürze": "apron", "Tank Top" : "tank top", "Crop-Top": "crop top",
+        "Pullover": "sweater", "Bluse": "blouse", "Hemd": "shirt",
+    }
+    
+    pose_keywords = {
+        "lacht": "smiling", "weint": "crying", "lächelt": "smiling",
+        "schläft": "sleeping", "schläfrig": "sleepy",
+        "sitzt": "sitting", "steht": "standing", "liegt": "lying down",
+        "fließt": "leaning forward", "geht": "walking",
+        "schaut": "looking at viewer", "dreht": "turning away",
+        "streckt": "stretching", "gähnt": "yawning",
+        "wütend": "angry", "traurig": "sad", "überrascht": "surprised",
+        "verlegen": "embarrassed", "errötet": "blushing",
+    }
+    
+    scene_keywords = {
+        "Café": "cafe interior", "Bar": "bar", "Küche": "kitchen",
+        "Schlafzimmer": "bedroom", "Badezimmer": "bathroom",
+        "Bett": "on bed", "Sofa": "on sofa", "Badewanne": "in bathtub",
+        "Draußen": "outdoors", "Park": "park", "Straße": "street",
+        "Regen": "rainy", "Sonne": "sunny", "Nacht": "night",
+    }
+    
+    found_outfits = set()
+    found_poses = set()
+    found_scenes = set()
+    
+    combined = " ".join(bot_msgs[-3:])  # Letzte 3 Bot-Nachrichten
+    
+    for de, en in outfit_keywords.items():
+        if de.lower() in combined.lower():
+            found_outfits.add(en)
+    for de, en in pose_keywords.items():
+        if de.lower() in combined.lower():
+            found_poses.add(en)
+    for de, en in scene_keywords.items():
+        if de.lower() in combined.lower():
+            found_scenes.add(en)
+    
+    parts = []
+    if found_outfits:
+        parts.append(f"wearing {', '.join(found_outfits)}")
+    if found_poses:
+        parts.append(", ".join(found_poses))
+    if found_scenes:
+        parts.append(", ".join(found_scenes))
+    
+    return ", ".join(parts) if parts else ""
+
+
 def build_prompt_from_appearance(character: dict, context: str = "") -> str:
     """Baut einen ComfyUI-Prompt aus dem Character-Aussehen."""
     app = character.get("appearance", {})
@@ -36,59 +107,116 @@ def build_prompt_from_appearance(character: dict, context: str = "") -> str:
     distinctive = app.get("distinctive", "")
     
     # Style prefix
+    # Style prefix and gender tags — adapt to style
     style_map = {
-        "anime": "anime style, anime art, ",
-        "realistic": "photorealistic, realistic, ",
-        "comic": "comic book style, western comic, ",
-        "cartoon": "cartoon style, ",
-        "pixel": "pixel art, ",
+        "anime": ("anime style, anime art, ", True),
+        "realistic": ("photorealistic, realistic, detailed skin texture, ", False),
+        "comic": ("comic book style, western comic, ", False),
+        "cartoon": ("cartoon style, ", False),
+        "pixel": ("pixel art, ", False),
     }
-    style_prefix = style_map.get(style, "anime style, ")
+    style_prefix, use_danbooru = style_map.get(style, ("anime style, ", True))
     
-    # Gender
+    # Gender — Danbooru tags for anime, natural language for others
+    if use_danbooru:
         if gender == "female":
             gender_tag = "1girl, woman, "
         elif gender == "male":
             gender_tag = "1boy, man, "
         else:
             gender_tag = "person, "
+    else:
+        if gender == "female":
+            gender_tag = "young woman, "
+        elif gender == "male":
+            gender_tag = "young man, "
+        else:
+            gender_tag = "person, "
+    
+    # Translation map: German appearance values → English tags for the model
+    translate_map = {
+        # Hair colors
+        "dunkelbraun": "dark brown", "hellbraun": "light brown", "braun": "brown",
+        "schwarz": "black", "blond": "blonde", "blonde": "blonde",
+        "rot": "red", "rotbraun": "auburn", "weiß": "white", "grau": "gray",
+        "blau": "blue", "grün": "green", "pink": "pink", "lila": "purple",
+        "silber": "silver", "feuerrot": "fiery red", "platinum": "platinum blonde",
+        # Hair styles
+        "kurz": "short", "kurz, ungeordnet": "short messy", "lang": "long",
+        "ungeordnet": "messy", "pony": "ponytail", "zöpfe": "twin braids",
+        "hochsteckfrisur": "updo", "gelockt": "curly", "glatt": "straight",
+        "wellig": "wavy", "kurz, gepflegt": "short neat", "schulterlang": "shoulder length",
+        # Eye colors
+        "braun": "brown", "blau": "blue", "grün": "green", "grau": "gray",
+        "bernstein": "amber", "haselnuss": "hazel", "violett": "violet",
+        # Skin
+        "hell": "fair", "dunkel": "dark", "mittel": "medium", "oliv": "olive",
+        "blass": "pale", "gebräunt": "tanned",
+        # Build
+        "schlank": "slim", "zierlich": "petite", "athletisch": "athletic",
+        "muskulös": "muscular", "kurvig": "curvy", "kräftig": "stocky",
+        "dünn": "thin", "groß": "tall", "klein": "short",
+        # Clothing keywords
+        "Café-Schürze über schwarzem T-Shirt": "cafe apron over black t-shirt",
+        "Café-Schürze über schwarzem T-Shirt, Jeans": "cafe apron over black t-shirt, jeans",
+        "Café-Schürze": "cafe apron", "Schürze": "apron", "T-Shirt": "t-shirt",
+        "Jeans": "jeans", "Kleid": "dress", "Anzug": "suit", "Hoodie": "hoodie",
+        "Jacke": "jacket", "Mantel": "coat", "Uniform": "uniform",
+        "über": "over", "schwarzem": "black", "weißem": "white",
+        "rotem": "red", "blauem": "blue", "grünem": "green",
+        # Distinctive features
+        "müde Augen": "tired eyes", "Augenringe": "dark circles under eyes",
+        "Kaugummi": "chewing gum", "Brille": "glasses", "Tattoo": "tattoo",
+        "Narben": "scars", "Sommersprossen": "freckles",
+        "leichte Augenringe": "slight dark circles under eyes",
+        "immer ein Kaugummi im Mund": "chewing gum",
+    }
+    
+    def tr(val: str) -> str:
+        """Übersetzt deutsche Wörter zu englischen Model-Tags."""
+        if not val:
+            return val
+        result = val
+        for de, en in sorted(translate_map.items(), key=lambda x: -len(x[0])):
+            result = result.replace(de, en)
+        return result
     
     # Build prompt
     parts = [
-        f"portrait, headshot, {style_prefix}{gender_tag}",
+        f"upper body, {style_prefix}{gender_tag}",
         f"{age} years old",
     ]
     if build:
-        parts.append(build)
+        parts.append(tr(build))
     if hair_color and hair_style:
-        parts.append(f"{hair_color} hair, {hair_style}")
+        parts.append(f"{tr(hair_color)} hair, {tr(hair_style)}")
     elif hair_color:
-        parts.append(f"{hair_color} hair")
+        parts.append(f"{tr(hair_color)} hair")
     elif hair_style:
-        parts.append(hair_style)
+        parts.append(tr(hair_style))
     if eye_color:
-        parts.append(f"{eye_color} eyes")
+        parts.append(f"{tr(eye_color)} eyes")
     if skin:
-        parts.append(f"{skin} skin")
+        parts.append(f"{tr(skin)} skin")
     if clothing:
-        parts.append(f"wearing {clothing}")
+        parts.append(f"wearing {tr(clothing)}")
     if distinctive:
-        parts.append(distinctive)
+        parts.append(tr(distinctive))
     if context:
         parts.append(context)
     
     prompt = ", ".join(parts)
-    prompt += ", detailed face, high quality, masterpiece, best quality, simple background"
+    prompt += ", looking at viewer, detailed face, high quality, masterpiece, best quality, simple background"
     
     return prompt
 
-def build_portrait_workflow(prompt: str, negative: str = "", width: int = 512, height: int = 512) -> dict:
+def build_portrait_workflow(prompt: str, negative: str = "", width: int = 1024, height: int = 1024) -> dict:
     return {
         "3": {
             "class_type": "KSampler",
             "inputs": {
                 "seed": int(time.time()) % (2**32),
-                "steps": 25,
+                "steps": 30,
                 "cfg": 7.0,
                 "sampler_name": "dpmpp_2m",
                 "scheduler": "karras",
@@ -158,7 +286,7 @@ def wait_for_image(prompt_id: str, timeout: int = 180) -> str | None:
         time.sleep(3)
     return None
 
-def build_i2i_workflow(prompt: str, negative: str, reference_image_path: str, width: int = 512, height: int = 512, denoise: float = 0.45) -> dict:
+def build_i2i_workflow(prompt: str, negative: str, reference_image_path: str, width: int = 1024, height: int = 1024, denoise: float = 0.45) -> dict:
     """Image-to-Image Workflow — nutzt ein Referenzbild für Konsistenz."""
     return {
         "3": {
@@ -241,7 +369,13 @@ def generate_portrait(character_name: str, character: dict, context: str = "") -
     AVATARS_DIR.mkdir(parents=True, exist_ok=True)
     
     prompt = build_prompt_from_appearance(character, context)
-    negative = "realistic, 3d, render"
+    style = character.get("appearance", {}).get("style", "anime")
+    # Negative prompt depends on style
+    if style == "anime":
+        negative = "realistic, 3d, render, photorealistic, "
+    else:
+        negative = "anime, 3d, render, "
+    negative += "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, blurry, deformed, ugly"
     
     # Check if we have a reference image for consistency
     safe_name = character_name.lower().replace(" ", "_")
@@ -249,15 +383,17 @@ def generate_portrait(character_name: str, character: dict, context: str = "") -
     
     if ref_image.exists():
         # Image-to-Image for consistency
+        # Higher denoise if context mentions outfit change
+        denoise = 0.55 if any(kw in (context or "").lower() for kw in ["outfit change", "wearing ", "nude", "lingerie", "bikini", "dress", "hoodie", "coat", "jacket"]) else 0.45
         try:
             uploaded_name = upload_image_to_comfy(str(ref_image))
-            workflow = build_i2i_workflow(prompt, negative, uploaded_name, 512, 512, denoise=0.45)
+            workflow = build_i2i_workflow(prompt, negative, uploaded_name, 1024, 1024, denoise=denoise)
         except Exception as e:
             print(f"I2I upload failed, falling back to T2I: {e}")
-            workflow = build_portrait_workflow(prompt, negative, 512, 512)
+            workflow = build_portrait_workflow(prompt, negative, 1024, 1024)
     else:
         # First generation — Text-to-Image
-        workflow = build_portrait_workflow(prompt, negative, 512, 512)
+        workflow = build_portrait_workflow(prompt, negative, 1024, 1024)
     
     try:
         prompt_id = queue_prompt(workflow)

main.py

@@ -15,7 +15,7 @@ from fastapi.staticfiles import StaticFiles
 from fastapi.templating import Jinja2Templates
 from pydantic import BaseModel
 
-from comfyui_integration import is_comfyui_running, generate_portrait
+from comfyui_integration import is_comfyui_running, generate_portrait, build_context_from_messages
 
 # === Konfiguration ===
 BASE_DIR = Path(__file__).parent
@@ -553,6 +553,17 @@ async def api_generate_portrait(name: str, request: Request):
     body = await request.json() if request.headers.get("content-type") == "application/json" else {}
     custom_prompt = body.get("prompt", "")
     
+    # Lese letzte Chat-Nachrichten für Kontext (Outfit-Änderungen etc.)
+    session_id = request.cookies.get(SESSION_COOKIE_NAME, "")
+    try:
+        if session_id:
+            recent_msgs = get_short_term_memory(session_id, name, limit=10)
+            chat_context = build_context_from_messages(recent_msgs, name)
+            if chat_context:
+                custom_prompt = f"{custom_prompt}, {chat_context}" if custom_prompt else chat_context
+    except Exception as e:
+        print(f"Context extraction failed (non-fatal): {e}")
+    
     import uuid
     job_id = str(uuid.uuid4())[:8]
     _portrait_jobs[job_id] = {"status": "generating", "image": None, "error": None}