michael/AI-Profile-Router
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add local FLUX image editing

Browse Source
This commit is contained in:
Mikei386
2026-08-30 10:07:40 +02:00
parent b18bc8964a
commit 6af25cf30c
13 changed files with 209 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files
router/ai_profile_router.py
+95 -17
View File
@@ -18,8 +18,9 @@ Virtuelle Modelle: qwen-fast, qwen-medium, qwen-large, qwen-ultra,
Kommandos: POST /fast, /medium, /large, /ultra, /uncensored
GET /status (Zustand)
Bildgenerierung (Z-Image-Turbo):
Bildgenerierung und Editing (FLUX.2-klein-4B):
POST /v1/images/generations (OpenAI-kompatibel)
POST /v1/images/edits (lokal, Referenzbilder)
GET /images (Liste)
GET /images/<datei> (PNG-Download)
@@ -37,7 +38,7 @@ Der Router leitet /v1/audio/speech und /v1/audio/transcriptions
per HTTP an die Worker weiter.
Der Router agiert als Modell-Orchestrator: vor der Generierung wird
llama.cpp gestoppt, der Bild-Worker lädt Z-Image, generiert und entlädt
llama.cpp gestoppt, der Bild-Worker lädt FLUX.2, generiert/bearbeitet und entlädt
das Modell wieder; danach wird das vorherige Qwen-Profil wiederher-
gestellt und erst dann geantwortet (try/finally – Qwen wird auch bei
Fehlgeschlagener Generierung wiederhergestellt).
@@ -116,7 +117,7 @@ CONNECT_TIMEOUT = float(os.environ.get("CONNECT_TIMEOUT", "10")) # s, Connect
POLL_INTERVAL = float(os.environ.get("POLL_INTERVAL", "2")) # s, Polling-Intervall
MAX_GENERATION_TOKENS = int(os.environ.get("MAX_GENERATION_TOKENS", "8192"))
# --- Bildgenerierung (Z-Image-Turbo) ---
# --- Bildgenerierung und Referenzbild-Bearbeitung (FLUX.2 Klein 4B) ---
LLAMA_SERVICE = os.environ.get("LLAMA_SERVICE", "mike-ai-llama-ui.service")
SYSTEMCTL_BIN = os.environ.get("SYSTEMCTL_BIN", "systemctl")
IMAGE_WORKER = os.environ.get(
@@ -125,7 +126,7 @@ IMAGE_PYTHON = os.environ.get(
"IMAGE_PYTHON", "/opt/mike-ai/ai-profile-router/venv/bin/python")
IMAGE_WORKER_URL = os.environ.get("IMAGE_WORKER_URL", "").rstrip("/")
IMAGE_WORKER_TOKEN = os.environ.get("IMAGE_WORKER_TOKEN", "").strip()
IMAGE_MODEL_NAME = os.environ.get("IMAGE_MODEL_NAME", "Z-Image-Turbo")
IMAGE_MODEL_NAME = os.environ.get("IMAGE_MODEL_NAME", "FLUX.2-klein-4B")
IMAGE_DIR = os.environ.get(
"IMAGE_DIR", "/opt/mike-ai/ai-profile-router/images")
IMAGE_WORKER_LOG = os.environ.get(
@@ -153,8 +154,8 @@ IMAGE_SIZES = {
"1920x1088": (1920, 1088),
"1088x1920": (1088, 1920),
}
# Z-Image-Turbo nutzt neun Scheduler-Schritte (acht DiT-Forwards) ohne CFG.
IMAGE_QUALITY = {"standard": 9, "high": 9}
# Das destillierte FLUX.2-klein-4B ist auf vier Schritte ausgelegt.
IMAGE_QUALITY = {"standard": 4, "high": 4}
IMAGE_DEFAULT_QUALITY = "standard"
IMAGE_MAX_N = 4
@@ -742,7 +743,7 @@ def switch_profile(profile: str, implicit: bool = False) -> None:
# ---------------------------------------------------------------------------
# Bildgenerierung (Z-Image-Turbo)
# Bildgenerierung und Editing (FLUX.2-klein-4B)
# ---------------------------------------------------------------------------
class _Worker:
@@ -986,7 +987,8 @@ def _restore_qwen(profile: str) -> None:
def generate_image(prompt: str, width: int, height: int, steps: int,
guidance: float, seed: int | None, n: int,
quality: str = "standard"
quality: str = "standard",
source_files: list[str] | None = None,
) -> tuple[list[str], str | None]:
"""Orchestriert die Bildgenerierung inkl. Qwen-Hotswap.
@@ -1047,6 +1049,7 @@ def generate_image(prompt: str, width: int, height: int, steps: int,
"guidance": guidance,
"seed": seed,
"output": output,
"source_files": source_files or [],
}, timeout=IMAGE_GEN_TIMEOUT)
if resp.get("status") != "ok":
raise RuntimeError(
@@ -1065,6 +1068,8 @@ def generate_image(prompt: str, width: int, height: int, steps: int,
"steps": steps,
"guidance": guidance,
"quality": quality,
"mode": "image-edit" if source_files else "text-to-image",
"reference_images": len(source_files or []),
"seconds": resp.get("seconds"),
"model": IMAGE_MODEL_NAME,
"created": time.strftime("%Y-%m-%dT%H:%M:%S"),
@@ -1403,6 +1408,12 @@ class Handler(BaseHTTPRequestHandler):
else:
self._send_error(503, "Bildgenerierung ist nicht installiert",
"server_error", "feature_disabled")
elif path == "/v1/images/edits" and self.command == "POST":
if ENABLE_IMAGE_GENERATION:
self._image_edit()
else:
self._send_error(503, "Bildbearbeitung ist nicht installiert",
"server_error", "feature_disabled")
elif path == "/v1/audio/speech" and self.command == "POST":
if ENABLE_TTS:
self._speech()
@@ -1627,22 +1638,89 @@ class Handler(BaseHTTPRequestHandler):
# ---------- Bildgenerierung ----------
def _image_generate(self) -> None:
data = self._read_image_request()
if data is not None:
self._image_request(data, [])
def _image_edit(self) -> None:
"""Edit with local image bytes supplied by the private Hermes plugin."""
data = self._read_image_request()
if data is None:
return
encoded: list[str] = []
primary = data.pop("image_b64", None)
if isinstance(primary, str) and primary:
encoded.append(primary)
references = data.pop("reference_images_b64", [])
if references is None:
references = []
if not isinstance(references, list) or any(
not isinstance(item, str) for item in references):
self._send_error(400, "'reference_images_b64' muss eine Liste sein",
"invalid_request_error", "invalid_references")
return
encoded.extend(references)
if not encoded:
self._send_error(400, "Referenzbild fehlt",
"invalid_request_error", "missing_image")
return
if len(encoded) > 4:
self._send_error(400, "höchstens vier Referenzbilder erlaubt",
"invalid_request_error", "too_many_images")
return
source_files: list[str] = []
try:
for item in encoded:
if item.startswith("data:"):
header, separator, item = item.partition(",")
if not separator or not header.lower().startswith("data:image/"):
raise ValueError("ungültige Bild-Data-URI")
try:
raw = base64.b64decode(item, validate=True)
except Exception as exc:
raise ValueError("ungültige Base64-Bilddaten") from exc
if not raw or len(raw) > CHAT_IMAGE_MAX_BYTES:
raise ValueError(
f"Referenzbild muss 1..{CHAT_IMAGE_MAX_BYTES} Bytes groß sein")
name = f".edit-{os.urandom(12).hex()}.ref"
os.makedirs(IMAGE_DIR, exist_ok=True)
with open(os.path.join(IMAGE_DIR, name), "xb") as output:
output.write(raw)
source_files.append(name)
self._image_request(data, source_files)
except ValueError as exc:
self._send_error(400, str(exc),
"invalid_request_error", "invalid_image")
finally:
for name in source_files:
try:
os.unlink(os.path.join(IMAGE_DIR, name))
except FileNotFoundError:
pass
except OSError as exc:
log.warning("temporäres Referenzbild nicht gelöscht: %s", exc)
def _read_image_request(self) -> dict | None:
try:
body = self._read_body()
except ValueError as e:
self._send_error(400, str(e),
"invalid_request_error", "invalid_body")
return
return None
try:
data = json.loads(body)
except ValueError:
self._send_error(400, "ungültiges JSON",
"invalid_request_error", "invalid_json")
return
return None
if not isinstance(data, dict):
self._send_error(400, "Request muss ein JSON-Objekt sein",
"invalid_request_error", "invalid_request")
return
return None
return data
def _image_request(self, data: dict, source_files: list[str]) -> None:
prompt = data.get("prompt")
if not isinstance(prompt, str) or not prompt.strip():
@@ -1679,19 +1757,19 @@ class Handler(BaseHTTPRequestHandler):
"invalid_request_error", "invalid_quality")
return
steps = data.get("steps", IMAGE_QUALITY[quality])
if not isinstance(steps, int) or isinstance(steps, bool) or steps != 9:
self._send_error(400, "Z-Image-Turbo erfordert 'steps'=9",
if not isinstance(steps, int) or isinstance(steps, bool) or steps != 4:
self._send_error(400, "FLUX.2-klein-4B erfordert 'steps'=4",
"invalid_request_error", "invalid_steps")
return
guidance = data.get("guidance", 0.0)
guidance = data.get("guidance", 1.0)
try:
guidance = float(guidance)
except (TypeError, ValueError):
self._send_error(400, "'guidance' muss eine Zahl sein",
"invalid_request_error", "invalid_guidance")
return
if guidance != 0.0:
self._send_error(400, "Z-Image-Turbo erfordert 'guidance'=0.0",
if guidance != 1.0:
self._send_error(400, "FLUX.2-klein-4B erfordert 'guidance'=1.0",
"invalid_request_error", "invalid_guidance")
return
@@ -1719,7 +1797,7 @@ class Handler(BaseHTTPRequestHandler):
try:
results, warning = generate_image(
prompt.strip(), width, height, steps, guidance, seed, n,
quality)
quality, source_files)
except (ValueError, RuntimeError) as e:
self._send_error(503, str(e), "server_error", "image_generation_failed")
return
router/image_worker.py
+20 -6
View File
@@ -1,5 +1,5 @@
#!/usr/bin/env python3
"""FLUX.2 [klein] 4B Base – Bild-Worker.
"""FLUX.2 [klein] 4B – Bild-Worker with reference-image editing.
Protokoll: zeilenbasiertes JSON über stdin/stdout.
@@ -32,7 +32,7 @@ os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
MODEL_DIR = os.environ.get(
"FLUX_MODEL_DIR", "/opt/mike-ai/models/FLUX.2-klein-base-4B")
"FLUX_MODEL_DIR", "/opt/mike-ai/models/FLUX.2-klein-4B")
_pipe = None # geladene Pipeline (None = entladen)
_load_seconds = 0.0 # Dauer des letzten Ladens
@@ -83,12 +83,13 @@ def _unload() -> None:
def _generate(req: dict) -> dict:
import torch
from PIL import Image
prompt = req["prompt"]
width = int(req.get("width", 1024))
height = int(req.get("height", 1024))
steps = int(req.get("steps", 50))
guidance = float(req.get("guidance", 4.0))
steps = int(req.get("steps", 4))
guidance = float(req.get("guidance", 1.0))
seed = req.get("seed")
output = req["output"]
@@ -98,14 +99,27 @@ def _generate(req: dict) -> dict:
generator = None
if seed is not None:
generator = torch.Generator(device="cuda").manual_seed(int(seed))
image = _pipe(
kwargs = dict(
prompt=prompt,
height=height,
width=width,
guidance_scale=guidance,
num_inference_steps=steps,
generator=generator,
).images[0]
)
source_files = req.get("source_files") or []
if not isinstance(source_files, list) or len(source_files) > 4:
raise ValueError("invalid source image list")
sources = []
for source in source_files:
if not isinstance(source, str):
raise ValueError("invalid source image filename")
path = os.path.join(os.path.dirname(output), source)
with Image.open(path) as opened:
sources.append(opened.convert("RGB"))
if sources:
kwargs["image"] = sources[0] if len(sources) == 1 else sources
image = _pipe(**kwargs).images[0]
os.makedirs(os.path.dirname(output) or ".", exist_ok=True)
image.save(output)