Add local FLUX image editing

.env.example

@@ -3,7 +3,7 @@ AI_BIND_ADDRESS=10.77.0.2
 MODEL_DIR=/data/models
 ROUTER_API_KEY=GENERATED_BY_INSTALLER
 CONTROLLER_TOKEN=GENERATED_BY_INSTALLER
-Z_IMAGE_MODEL_DIR=/data/models/Z-Image-Turbo
+FLUX_MODEL_DIR=/data/models/FLUX.2-klein-4B
 IMAGE_GPU_DEVICES=1
 PIPER_TTS_VERSION=1.6.0
 PIPER_VOICE=de_DE-thorsten-high

ATHENA.md

@@ -10,7 +10,7 @@ Sie betreibt:
 
 - llama.cpp mit genau einem aktiven Qwen-Profil,
 - den OpenAI-kompatiblen Profile Router,
-- Z-Image-Turbo für Bilder,
+- FLUX.2-klein-4B für Textbilder und Referenzbild-Bearbeitung,
 - XTTS und Piper für Sprache,
 - das Athena-Dashboard,
 - WireGuard-Gateway und Datenbackup,
@@ -48,7 +48,7 @@ Qwen-Profil wird vom Profile Controller verwaltet.
 - Fast: kurze, interaktive Aufgaben
 - Medium/Large/Ultra: steigende Kontextgrößen desselben lokalen Qwen-Modells
 - Uncensored: separates lokales Profil
-- Z-Image-Turbo: Bildgenerierung; Qwen wird dafür kurz entladen und danach
+- FLUX.2-klein-4B: Bildgenerierung und Editing; Qwen wird dafür kurz entladen und danach
   automatisch wiederhergestellt
 - XTTS: RTX 3060; Piper bleibt CPU-Fallback
 

README.md

@@ -10,7 +10,7 @@ Bild- und Sprachausgabe. **Hermes und die Fach-MCPs laufen auf Unraid.**
 
 - genau ein aktives llama.cpp-Profil: Fast, Medium, Large, Ultra oder Uncensored
 - Profile Router auf Port 8081
-- Z-Image-Turbo als exklusiver Bild-Worker auf der RTX 5080
+- FLUX.2-klein-4B für Textbilder und Referenzbild-Bearbeitung auf der RTX 5080
 - XTTS auf der RTX 3060 mit Piper als CPU-Fallback
 - Live-Dashboard mit 21 Tagen Detailhistorie auf Port 8099
 - WireGuard-Gateway, Datenbackup und Athena-Operator

compose.yaml

@@ -573,7 +573,7 @@ services:
       IMAGE_DIR: /data/images
       IMAGE_WORKER_URL: http://image-worker:8086
       IMAGE_WORKER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"
-      IMAGE_MODEL_NAME: Z-Image-Turbo
+      IMAGE_MODEL_NAME: FLUX.2-klein-4B
       CHAT_IMAGE_ALLOW_REMOTE_URLS: "false"
       ENABLE_IMAGE_GENERATION: "true"
       ENABLE_TTS: "true"
@@ -628,13 +628,13 @@ services:
     read_only: true
     tmpfs: ["/tmp:size=1g,mode=1777"]
     volumes:
-      - "${Z_IMAGE_MODEL_DIR:-/data/models/Z-Image-Turbo}:/models/Z-Image-Turbo:ro"
+      - "${FLUX_MODEL_DIR:-/data/models/FLUX.2-klein-4B}:/models/FLUX.2-klein-4B:ro"
       - router-images:/data/images
     environment:
       NVIDIA_VISIBLE_DEVICES: ${IMAGE_GPU_DEVICES:-1}
       NVIDIA_DRIVER_CAPABILITIES: compute,utility
       WORKER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"
-      Z_IMAGE_MODEL_DIR: /models/Z-Image-Turbo
+      FLUX_MODEL_DIR: /models/FLUX.2-klein-4B
       IMAGE_DIR: /data/images
     networks: [inference]
     security_opt: ["no-new-privileges:true"]

config/install.env.example

@@ -16,7 +16,7 @@ NVIDIA_MIN_DRIVER_MAJOR=570
 TEXT_GPU_DEVICES=0
 SECONDARY_GPU_DEVICES=1
 IMAGE_GPU_DEVICES=1
-Z_IMAGE_MODEL_DIR=/data/models/Z-Image-Turbo
+FLUX_MODEL_DIR=/data/models/FLUX.2-klein-4B
 
 # Headless remote reachability. Firmware power-loss recovery is configured
 # separately once at the physical machine.

deploy/install.sh

@@ -15,7 +15,7 @@ OLD_SERVICE=mike-ai-local-llm-router.service
 OLD_DIR=/opt/mike-ai/local-llm-router
 BACKUP_DIR=/opt/mike-ai/.backup-ai-profile-router-$(date +%Y%m%d-%H%M%S)
 VENV="$INSTALL_DIR/venv"
-MODEL_DIR=/opt/mike-ai/models/FLUX.2-klein-base-4B
+MODEL_DIR=/opt/mike-ai/models/FLUX.2-klein-4B
 IMAGE_DIR="$INSTALL_DIR/images"
 XTTS_VENV=/opt/mike-ai/xtts/venv
 XTTS_CACHE=/opt/mike-ai/xtts/.cache
@@ -72,13 +72,14 @@ echo "-- Installiere festgeschriebene Bild-Abhängigkeiten"
 if [ -f "$MODEL_DIR/model_index.json" ]; then
   echo "-- FLUX-Modell vorhanden: $MODEL_DIR"
 else
-  echo "-- Lade FLUX.2-klein-base-4B nach $MODEL_DIR (kann dauern)"
+  echo "-- Lade FLUX.2-klein-4B nach $MODEL_DIR (kann dauern)"
   "$VENV/bin/python" - <<'PY'
 import os
 from huggingface_hub import snapshot_download
 snapshot_download(
-    repo_id="black-forest-labs/FLUX.2-klein-base-4B",
+    repo_id="black-forest-labs/FLUX.2-klein-4B",
-    local_dir="/opt/mike-ai/models/FLUX.2-klein-base-4B",
+    revision="e7b7dc27f91deacad38e78976d1f2b499d76a294",
+    local_dir="/opt/mike-ai/models/FLUX.2-klein-4B",
     local_dir_use_symlinks=False,
 )
 print("Modell-Download abgeschlossen")

dev/test_local.sh

@@ -343,6 +343,33 @@ CTYPE=$(grep -i content-type /tmp/hdr12.txt | tr -d "\r")
 [ "$CODE" = "200" ] && [ -s /tmp/test_dl.png ] && echo "$CTYPE" | grep -qi "image/png" \
   && ok "PNG-Download (200, $CTYPE)" || bad "PNG-Download (Code $CODE, $CTYPE)"
 
+# --- 12b. Referenzbild-Bearbeitung ------------------------------------------------------
+echo "== Test 12b: POST /v1/images/edits mit lokalem Referenzbild"
+python3 - <<'PY' >/tmp/edit-request.json
+import base64, json
+png = open('/tmp/test_dl.png', 'rb').read()
+print(json.dumps({
+    'prompt': 'Behalte die Person bei und ändere nur den Hintergrund',
+    'size': '1024x1024',
+    'steps': 4,
+    'guidance': 1.0,
+    'response_format': 'b64_json',
+    'image_b64': base64.b64encode(png).decode(),
+}))
+PY
+rm -f /tmp/test_worker_requests.jsonl
+RESP=$(curl -sf "$BASE/v1/images/edits" -H "Content-Type: application/json" \
+  --data-binary @/tmp/edit-request.json)
+echo "$RESP" | python3 -c '
+import base64,json,sys
+d=json.load(sys.stdin)
+assert base64.b64decode(d["data"][0]["b64_json"])[:4] == b"\x89PNG"
+' || bad "Bildbearbeitung liefert kein PNG"
+REFS=$(tail -1 /tmp/test_worker_requests.jsonl | python3 -c \
+  'import json,sys; print(len(json.load(sys.stdin).get("source_files", [])))')
+[ "$REFS" = "1" ] && ! find /tmp/test-images -name '.edit-*.ref' | grep -q . \
+  && ok "Referenzbild übergeben und danach gelöscht" || bad "Referenzbild-Pfad/Cleanup"
+
 # --- 13. Bild-Liste ---------------------------------------------------------------------
 echo "== Test 13: GET /images"
 RESP=$(curl -sf "$BASE/images")

docs/ARCHITECTURE.md

@@ -6,7 +6,7 @@ flowchart LR
     H -->|OpenAI API| R[Profile Router<br/>Athena :8081]
     R --> P[Profile Controller]
     P --> Q[genau ein llama.cpp-Profil<br/>Qwen Fast / Medium / Large / Ultra / Uncensored]
-    R --> I[Z-Image-Turbo<br/>RTX 5080, bei Bedarf]
+    R --> I[FLUX.2-klein-4B<br/>RTX 5080, Text + Editing]
     R --> T[XTTS RTX 3060<br/>Piper CPU-Fallback]
 
     H --> U[MUA / Unraid MCP]
@@ -49,4 +49,4 @@ Kontextgröße:
 Die visuelle Fassung liegt als `athena-architecture-map.png` neben dieser Datei.
 Eine zweite Detailkarte, `athena-gpu-allocation-map.png`, zeigt die
 profilabhängige Layer-Verteilung auf RTX 5080 und RTX 3060 sowie die festen
-GPU-Zuordnungen von Z-Image, Vision-Projektor und XTTS.
+GPU-Zuordnungen von FLUX.2, Vision-Projektor und XTTS.

install.sh

@@ -341,7 +341,7 @@ UNCENSORED_TENSOR_SPLIT=${UNCENSORED_TENSOR_SPLIT:-90,10}
 UNCENSORED_MTP_MAX=${UNCENSORED_MTP_MAX:-2}
 EXPERIMENTAL_GPU_DEVICES=${TEXT_GPU_DEVICES:-0}
 IMAGE_GPU_DEVICES=${IMAGE_GPU_DEVICES:-${TEXT_GPU_DEVICES:-0}}
-Z_IMAGE_MODEL_DIR=${Z_IMAGE_MODEL_DIR:-/data/models/Z-Image-Turbo}
+FLUX_MODEL_DIR=${FLUX_MODEL_DIR:-/data/models/FLUX.2-klein-4B}
 LLAMA_THREADS=${LLAMA_THREADS:-6}
 LLAMA_THREADS_BATCH=${LLAMA_THREADS_BATCH:-6}
 EOF
@@ -460,14 +460,14 @@ build_and_start() {
   docker build --progress=plain --build-arg LLAMA_CPP_COMMIT="$commit" \
     -f platform/docker/llama-cpp/Dockerfile -t mike-ai/llama.cpp:local .
   docker compose --env-file "$SECRETS_DIR/stack.env" --profile image build image-worker
-  if [[ ! -s ${Z_IMAGE_MODEL_DIR:-/data/models/Z-Image-Turbo}/model_index.json ]]; then
+  if [[ ! -s ${FLUX_MODEL_DIR:-/data/models/FLUX.2-klein-4B}/model_index.json ]]; then
-    log "Z-Image-Turbo laden"
+    log "FLUX.2-klein-4B laden"
-    install -d -m 0755 "${Z_IMAGE_MODEL_DIR:-/data/models/Z-Image-Turbo}"
+    install -d -m 0755 "${FLUX_MODEL_DIR:-/data/models/FLUX.2-klein-4B}"
     docker run --rm --entrypoint python \
-      -v "${Z_IMAGE_MODEL_DIR:-/data/models/Z-Image-Turbo}:/download" \
+      -v "${FLUX_MODEL_DIR:-/data/models/FLUX.2-klein-4B}:/download" \
       mike-ai/image-worker:local -c \
-      "from huggingface_hub import snapshot_download; snapshot_download('Tongyi-MAI/Z-Image-Turbo', revision='f332072aa78be7aecdf3ee76d5c247082da564a6', local_dir='/download')"
+      "from huggingface_hub import snapshot_download; snapshot_download('black-forest-labs/FLUX.2-klein-4B', revision='e7b7dc27f91deacad38e78976d1f2b499d76a294', local_dir='/download')"
-    chmod -R a-w "${Z_IMAGE_MODEL_DIR:-/data/models/Z-Image-Turbo}"
+    chmod -R a-w "${FLUX_MODEL_DIR:-/data/models/FLUX.2-klein-4B}"
   fi
   # Creates the tools network and deploys the only host-bound MCP: Operator.
   # Portable MCPs and Hermes live on Unraid and are restored through Appdata.

platform/docker/image-worker/image_worker.py

@@ -1,5 +1,10 @@
 #!/usr/bin/env python3
-"""Private Z-Image-Turbo worker used only during a GPU hot swap."""
+"""Private FLUX.2 Klein 4B worker used only during a GPU hot swap.
+
+The same pipeline handles text-to-image and local reference-image editing.
+Reference images are exchanged with the router through the shared image
+volume; request bodies therefore never contain private image bytes here.
+"""
 
 from __future__ import annotations
 
@@ -14,7 +19,7 @@ from pathlib import Path
 HOST = os.environ.get("WORKER_HOST", "0.0.0.0")
 PORT = int(os.environ.get("WORKER_PORT", "8086"))
 TOKEN = os.environ.get("WORKER_TOKEN", "").strip()
-MODEL_DIR = os.environ.get("Z_IMAGE_MODEL_DIR", "/models/Z-Image-Turbo")
+MODEL_DIR = os.environ.get("FLUX_MODEL_DIR", "/models/FLUX.2-klein-4B")
 OUTPUT_DIR = Path(os.environ.get("IMAGE_DIR", "/data/images")).resolve()
 PIPE = None
 LOAD_SECONDS = 0.0
@@ -33,15 +38,13 @@ def load_pipeline() -> None:
     if PIPE is not None:
         return
     import torch
-    from diffusers import ZImagePipeline
+    from diffusers import Flux2KleinPipeline
     started = time.monotonic()
-    PIPE = ZImagePipeline.from_pretrained(
+    PIPE = Flux2KleinPipeline.from_pretrained(
         MODEL_DIR, torch_dtype=torch.bfloat16, low_cpu_mem_usage=False)
-    # The Qwen text encoder and the DiT do not fit together in the usable
+    # Officially supported low-VRAM path. It keeps the complete pipeline
-    # 16 GiB of the RTX 5080. Sequential offload keeps only the active
+    # within the usable 16 GiB of the RTX 5080 and leaves the RTX 3060 alone.
-    # submodule on CUDA. This is slower than a fully resident pipeline, but
+    PIPE.enable_model_cpu_offload()
-    # deterministic and leaves the RTX 3060 available for XTTS.
-    PIPE.enable_sequential_cpu_offload()
     if hasattr(PIPE, "enable_vae_slicing"):
         PIPE.enable_vae_slicing()
     if hasattr(PIPE, "enable_vae_tiling"):
@@ -51,6 +54,7 @@ def load_pipeline() -> None:
 
 def generate(data: dict) -> dict:
     import torch
+    from PIL import Image
     prompt = data.get("prompt")
     filename = data.get("filename")
     if not isinstance(prompt, str) or not prompt.strip() or len(prompt) > 8000:
@@ -62,17 +66,37 @@ def generate(data: dict) -> dict:
     if (width, height) not in {(1024, 1024), (1536, 1024), (1024, 1536),
                                (1920, 1088), (1088, 1920)}:
         raise ValueError("unsupported image size")
-    steps = int(data.get("steps", 9))
+    steps = int(data.get("steps", 4))
-    guidance = float(data.get("guidance", 0.0))
+    guidance = float(data.get("guidance", 1.0))
-    if steps != 9 or guidance != 0.0:
+    if steps != 4 or guidance != 1.0:
-        raise ValueError("Z-Image-Turbo requires steps=9 and guidance=0.0")
+        raise ValueError("FLUX.2-klein-4B requires steps=4 and guidance=1.0")
+    source_files = data.get("source_files") or []
+    if not isinstance(source_files, list) or len(source_files) > 4:
+        raise ValueError("invalid source image list")
+    source_images = []
+    for source_name in source_files:
+        if not isinstance(source_name, str) or Path(source_name).name != source_name:
+            raise ValueError("invalid source image filename")
+        source = (OUTPUT_DIR / source_name).resolve()
+        if source.parent != OUTPUT_DIR or not source.is_file():
+            raise ValueError("source image not found")
+        with Image.open(source) as opened:
+            source_images.append(opened.convert("RGB"))
     seed = data.get("seed")
     generator = None if seed is None else torch.Generator(device="cuda").manual_seed(int(seed))
     load_pipeline()
     started = time.monotonic()
-    image = PIPE(prompt=prompt, height=height, width=width,
+    kwargs = {
-                 num_inference_steps=9, guidance_scale=0.0,
+        "prompt": prompt,
-                 generator=generator).images[0]
+        "height": height,
+        "width": width,
+        "num_inference_steps": 4,
+        "guidance_scale": 1.0,
+        "generator": generator,
+    }
+    if source_images:
+        kwargs["image"] = source_images[0] if len(source_images) == 1 else source_images
+    image = PIPE(**kwargs).images[0]
     OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
     output = OUTPUT_DIR / filename
     image.save(output)
@@ -84,7 +108,7 @@ def generate(data: dict) -> dict:
 class Handler(BaseHTTPRequestHandler):
     def log_message(self, fmt: str, *args: object) -> None:
         # Never log request bodies/prompts.
-        print(f"[z-image-worker] {self.client_address[0]} {fmt % args}", flush=True)
+        print(f"[flux-image-worker] {self.client_address[0]} {fmt % args}", flush=True)
 
     def reply(self, status: int, payload: dict) -> None:
         body = json.dumps(payload, separators=(",", ":")).encode()
@@ -113,7 +137,7 @@ class Handler(BaseHTTPRequestHandler):
                 raise ValueError("invalid request size")
             self.reply(200, generate(json.loads(self.rfile.read(length))))
         except Exception as exc:
-            print(f"[z-image-worker] generation failed: "
+            print(f"[flux-image-worker] generation failed: "
                   f"{type(exc).__name__}: {str(exc)[:1000]}", flush=True)
             self.reply(400, {"status": "error", "message": str(exc)})
 

platform/models/manifest.example.yaml

@@ -28,8 +28,9 @@ models:
     sha256: "REPLACE_AFTER_VERIFICATION"
   image:
     role: image-generation
-    source: Tongyi-MAI/Z-Image-Turbo
+    source: black-forest-labs/FLUX.2-klein-4B
-    target: /data/models/Z-Image-Turbo
+    revision: e7b7dc27f91deacad38e78976d1f2b499d76a294
+    target: /data/models/FLUX.2-klein-4B
     revision: "f332072aa78be7aecdf3ee76d5c247082da564a6"
   xtts:
     role: text-to-speech

router/ai_profile_router.py

@@ -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.
+# Das destillierte FLUX.2-klein-4B ist auf vier Schritte ausgelegt.
-IMAGE_QUALITY = {"standard": 9, "high": 9}
+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:
+        if not isinstance(steps, int) or isinstance(steps, bool) or steps != 4:
-            self._send_error(400, "Z-Image-Turbo erfordert 'steps'=9",
+            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:
+        if guidance != 1.0:
-            self._send_error(400, "Z-Image-Turbo erfordert 'guidance'=0.0",
+            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

@@ -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))
+    steps = int(req.get("steps", 4))
-    guidance = float(req.get("guidance", 4.0))
+    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)