arch_agent
4d0c915b2a
- portraitAdded guard prevents duplicate image in chat - Negative prompt: multiple views, split screen, repetition, duplicate - Removed cowboy shot (caused repetition) - Clean gender tags: 1girl, solo (no trailing comma) - Steps 35, CFG 6.5 for better NoobAI-XL quality - Image max-width 400px in chat
417 lines
16 KiB
Python
417 lines
16 KiB
Python
"""
|
|
NeonChat ComfyUI Integration — Profilbild-Generierung für Characters
|
|
Basiert auf dem Character-Aussehen (appearance)
|
|
"""
|
|
|
|
import json, os, time, shutil
|
|
from pathlib import Path
|
|
import httpx
|
|
|
|
COMFYUI_URL = os.environ.get("COMFYUI_URL", "http://localhost:8188")
|
|
COMFYUI_OUTPUT_DIR = Path(os.environ.get("COMFYUI_OUTPUT", str(Path.home() / "comfy" / "ComfyUI" / "output")))
|
|
AVATARS_DIR = Path(__file__).parent / "static" / "avatars"
|
|
|
|
def is_comfyui_running() -> bool:
|
|
try:
|
|
with httpx.Client(timeout=3) as client:
|
|
resp = client.get(f"{COMFYUI_URL}/system_stats")
|
|
return resp.status_code == 200
|
|
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", {})
|
|
|
|
style = app.get("style", "anime")
|
|
gender = app.get("gender", "female")
|
|
age = app.get("age", "25")
|
|
height = app.get("height", "")
|
|
build = app.get("build", "")
|
|
hair_color = app.get("hair_color", "")
|
|
hair_style = app.get("hair_style", "")
|
|
eye_color = app.get("eye_color", "")
|
|
skin = app.get("skin", "")
|
|
clothing = app.get("clothing", "")
|
|
distinctive = app.get("distinctive", "")
|
|
|
|
# Style prefix
|
|
# Style prefix and gender tags — adapt to style
|
|
style_map = {
|
|
"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, use_danbooru = style_map.get(style, ("anime style, ", True))
|
|
|
|
# Gender — Danbooru tags for anime, natural language for others
|
|
if use_danbooru:
|
|
if gender == "female":
|
|
gender_tag = "1girl, solo"
|
|
elif gender == "male":
|
|
gender_tag = "1boy, solo"
|
|
else:
|
|
gender_tag = "solo"
|
|
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 — use established Danbooru tags for NoobAI-XL
|
|
parts = [
|
|
f"{style_prefix}{gender_tag}",
|
|
f"{age} years old",
|
|
]
|
|
if build:
|
|
parts.append(tr(build))
|
|
if hair_color and hair_style:
|
|
parts.append(f"{tr(hair_color)} hair, {tr(hair_style)}")
|
|
elif hair_color:
|
|
parts.append(f"{tr(hair_color)} hair")
|
|
elif hair_style:
|
|
parts.append(tr(hair_style))
|
|
if eye_color:
|
|
parts.append(f"{tr(eye_color)} eyes")
|
|
if skin:
|
|
parts.append(f"{tr(skin)} skin")
|
|
if clothing:
|
|
parts.append(f"wearing {tr(clothing)}")
|
|
if distinctive:
|
|
parts.append(tr(distinctive))
|
|
if context:
|
|
parts.append(context)
|
|
|
|
prompt = ", ".join(parts)
|
|
prompt += ", looking at viewer, detailed face, masterpiece, best quality, very aesthetic, absurdres"
|
|
|
|
return prompt
|
|
|
|
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": 35,
|
|
"cfg": 6.5,
|
|
"sampler_name": "dpmpp_2m",
|
|
"scheduler": "karras",
|
|
"denoise": 1.0,
|
|
"model": ["4", 0],
|
|
"positive": ["6", 0],
|
|
"negative": ["7", 0],
|
|
"latent_image": ["5", 0]
|
|
}
|
|
},
|
|
"4": {
|
|
"class_type": "CheckpointLoaderSimple",
|
|
"inputs": {"ckpt_name": "NoobAI-XL-v1.1.safetensors"}
|
|
},
|
|
"5": {
|
|
"class_type": "EmptyLatentImage",
|
|
"inputs": {"width": width, "height": height, "batch_size": 1}
|
|
},
|
|
"6": {
|
|
"class_type": "CLIPTextEncode",
|
|
"inputs": {
|
|
"text": prompt,
|
|
"clip": ["4", 1]
|
|
}
|
|
},
|
|
"7": {
|
|
"class_type": "CLIPTextEncode",
|
|
"inputs": {
|
|
"text": f"realistic, 3d, render, photorealistic, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, blurry, deformed, ugly, multiple views, split screen, repetition, duplicate, {negative}",
|
|
"clip": ["4", 1]
|
|
}
|
|
},
|
|
"8": {
|
|
"class_type": "VAEDecode",
|
|
"inputs": {"samples": ["3", 0], "vae": ["4", 2]}
|
|
},
|
|
"9": {
|
|
"class_type": "SaveImage",
|
|
"inputs": {"filename_prefix": "neonchat_portrait", "images": ["8", 0]}
|
|
}
|
|
}
|
|
|
|
def queue_prompt(workflow: dict) -> str:
|
|
with httpx.Client(timeout=10) as client:
|
|
resp = client.post(f"{COMFYUI_URL}/prompt", json={"prompt": workflow})
|
|
result = resp.json()
|
|
return result.get("prompt_id", "")
|
|
|
|
def get_history(prompt_id: str) -> dict:
|
|
try:
|
|
with httpx.Client(timeout=5) as client:
|
|
resp = client.get(f"{COMFYUI_URL}/history/{prompt_id}")
|
|
return resp.json()
|
|
except:
|
|
return {}
|
|
|
|
def wait_for_image(prompt_id: str, timeout: int = 180) -> str | None:
|
|
start = time.time()
|
|
while time.time() - start < timeout:
|
|
history = get_history(prompt_id)
|
|
if prompt_id in history:
|
|
outputs = history[prompt_id].get("outputs", {})
|
|
if "9" in outputs:
|
|
images = outputs["9"].get("images", [])
|
|
if images:
|
|
return images[0].get("filename", "")
|
|
time.sleep(3)
|
|
return None
|
|
|
|
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": {
|
|
"class_type": "KSampler",
|
|
"inputs": {
|
|
"seed": int(time.time()) % (2**32),
|
|
"steps": 25,
|
|
"cfg": 7.0,
|
|
"sampler_name": "dpmpp_2m",
|
|
"scheduler": "karras",
|
|
"denoise": denoise,
|
|
"model": ["4", 0],
|
|
"positive": ["6", 0],
|
|
"negative": ["7", 0],
|
|
"latent_image": ["10", 0]
|
|
}
|
|
},
|
|
"4": {
|
|
"class_type": "CheckpointLoaderSimple",
|
|
"inputs": {"ckpt_name": "NoobAI-XL-v1.1.safetensors"}
|
|
},
|
|
"5": {
|
|
"class_type": "LoadImage",
|
|
"inputs": {"image": reference_image_path}
|
|
},
|
|
"6": {
|
|
"class_type": "CLIPTextEncode",
|
|
"inputs": {
|
|
"text": prompt,
|
|
"clip": ["4", 1]
|
|
}
|
|
},
|
|
"7": {
|
|
"class_type": "CLIPTextEncode",
|
|
"inputs": {
|
|
"text": f"realistic, 3d, render, photorealistic, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, blurry, deformed, ugly, multiple views, split screen, repetition, duplicate, {negative}",
|
|
"clip": ["4", 1]
|
|
}
|
|
},
|
|
"8": {
|
|
"class_type": "VAEDecode",
|
|
"inputs": {"samples": ["3", 0], "vae": ["4", 2]}
|
|
},
|
|
"9": {
|
|
"class_type": "SaveImage",
|
|
"inputs": {"filename_prefix": "neonchat_portrait", "images": ["8", 0]}
|
|
},
|
|
"10": {
|
|
"class_type": "VAEEncode",
|
|
"inputs": {"pixels": ["11", 0], "vae": ["4", 2]}
|
|
},
|
|
"11": {
|
|
"class_type": "ImageScale",
|
|
"inputs": {
|
|
"image": ["5", 0],
|
|
"upscale_method": "bilinear",
|
|
"width": width,
|
|
"height": height,
|
|
"crop": "center"
|
|
}
|
|
}
|
|
}
|
|
|
|
def upload_image_to_comfy(image_path: str) -> str:
|
|
"""Lädt ein Bild zu ComfyUI hoch und gibt den Dateinamen zurück."""
|
|
import os
|
|
filename = os.path.basename(image_path)
|
|
with open(image_path, "rb") as f:
|
|
files = {"image": (filename, f, "image/png")}
|
|
with httpx.Client(timeout=30) as client:
|
|
resp = client.post(f"{COMFYUI_URL}/upload/image", files=files)
|
|
data = resp.json()
|
|
return data.get("name", filename)
|
|
|
|
def generate_portrait(character_name: str, character: dict, context: str = "") -> str | None:
|
|
"""Generiert ein Profilbild. Nutzt Image-to-Image wenn ein Referenzbild existiert."""
|
|
if not is_comfyui_running():
|
|
return None
|
|
|
|
AVATARS_DIR.mkdir(parents=True, exist_ok=True)
|
|
|
|
prompt = build_prompt_from_appearance(character, context)
|
|
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(" ", "_")
|
|
ref_image = AVATARS_DIR / f"{safe_name}.png"
|
|
|
|
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, 1024, 1024, denoise=denoise)
|
|
except Exception as e:
|
|
print(f"I2I upload failed, falling back to T2I: {e}")
|
|
workflow = build_portrait_workflow(prompt, negative, 1024, 1024)
|
|
else:
|
|
# First generation — Text-to-Image
|
|
workflow = build_portrait_workflow(prompt, negative, 1024, 1024)
|
|
|
|
try:
|
|
prompt_id = queue_prompt(workflow)
|
|
if not prompt_id:
|
|
return None
|
|
|
|
filename = wait_for_image(prompt_id, timeout=180)
|
|
if not filename:
|
|
return None
|
|
|
|
comfy_output = COMFYUI_OUTPUT_DIR / filename
|
|
if not comfy_output.exists():
|
|
return None
|
|
|
|
dest = AVATARS_DIR / f"{safe_name}.png"
|
|
shutil.copy2(str(comfy_output), str(dest))
|
|
|
|
return f"/static/avatars/{safe_name}.png"
|
|
except Exception as e:
|
|
print(f"ComfyUI error: {e}")
|
|
return None |