Integrate Qwen image prompt enhancers

compose.yaml

@@ -572,6 +572,8 @@ services:
       ALLOWED_PROFILES: fast,medium,large,ultra,uncensored
       IMAGE_WORKER: image
       RESTORE_WORKER: restore
+      IMAGE_PROMPT_I2I_WORKER: image-prompt-i2i
+      IMAGE_PROMPT_T2I_WORKER: image-prompt-t2i
       FLUX_STANDBY_WORKER: flux-standby
       TTS_WORKER: qwen3
       MUSIC_WORKER: acestep
@@ -602,6 +604,8 @@ services:
     volumes:
       - ./router/router_profiles.json:/etc/mike-ai/router-profiles.json:ro
       - ./config/global-system-policy.txt:/etc/mike-ai/global-system-policy.txt:ro
+      - "${QWEN_IMAGE_PE_I2I_MODEL_DIR:-/data/models/qwen-image-2.1-pe-i2i-q5}/system_prompt.txt:/etc/mike-ai/qwen-image-pe-i2i-system-prompt.txt:ro"
+      - "${QWEN_IMAGE_PE_T2I_MODEL_DIR:-/data/models/qwen-image-2.1-pe-t2i-q5}/system_prompt.txt:/etc/mike-ai/qwen-image-pe-t2i-system-prompt.txt:ro"
       - router-state:/var/lib/mike-ai-profile-router
       - router-images:/data/images
     environment:
@@ -631,6 +635,11 @@ services:
       IMAGE_DIR: /data/images
       IMAGE_WORKER_URL: http://image-worker:8086
       IMAGE_WORKER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"
+      IMAGE_PROMPT_I2I_URL: http://image-prompt-enhancer-i2i:8080
+      IMAGE_PROMPT_T2I_URL: http://image-prompt-enhancer-t2i:8080
+      IMAGE_PROMPT_I2I_SYSTEM_FILE: /etc/mike-ai/qwen-image-pe-i2i-system-prompt.txt
+      IMAGE_PROMPT_T2I_SYSTEM_FILE: /etc/mike-ai/qwen-image-pe-t2i-system-prompt.txt
+      IMAGE_PROMPT_ENHANCER_TIMEOUT: "180"
       IMAGE_MODEL_NAME: Qwen-Image-2.1-int8
       IMAGE_INFERENCE_STEPS: "25"
       CHAT_IMAGE_ALLOW_REMOTE_URLS: "false"
@@ -716,6 +725,143 @@ services:
       retries: 90
       start_period: 10s
 
+  image-prompt-enhancer-i2i:
+    image: mike-ai/llama.cpp:b10930
+    container_name: mike-ai-image-prompt-enhancer-i2i
+    restart: "no"
+    profiles: [image]
+    labels:
+      com.mike-ai.image-worker: image-prompt-i2i
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ["${QWEN_IMAGE_PE_GPU:-0}"]
+              capabilities: [gpu]
+    read_only: true
+    tmpfs: ["/tmp:size=512m,mode=1777"]
+    volumes:
+      - "${QWEN_IMAGE_PE_I2I_MODEL_DIR:-/data/models/qwen-image-2.1-pe-i2i-q5}:/models:ro"
+    command:
+      - --model
+      - /models/Qwen-Image-2.1-PE-I2I.Q5_K_M.gguf
+      - --mmproj
+      - /models/Qwen-Image-2.1-PE-I2I.mmproj-bf16.gguf
+      - --mmproj-offload
+      - --mmproj-device
+      - CUDA0
+      - --alias
+      - qwen-image-pe-i2i
+      - --ctx-size
+      - "16384"
+      - --flash-attn
+      - "on"
+      - --cache-type-k
+      - q8_0
+      - --cache-type-v
+      - q8_0
+      - --threads
+      - "6"
+      - --threads-batch
+      - "6"
+      - --batch-size
+      - "1024"
+      - --ubatch-size
+      - "128"
+      - --parallel
+      - "1"
+      - --jinja
+      - --reasoning
+      - auto
+      - --host
+      - 0.0.0.0
+      - --port
+      - "8080"
+      - --metrics
+      - --n-gpu-layers
+      - all
+      - --device
+      - CUDA0
+      - --split-mode
+      - none
+      - --no-ui
+    networks: [inference]
+    security_opt: ["no-new-privileges:true"]
+    cap_drop: [ALL]
+    healthcheck:
+      test: [CMD, curl, -fsS, "http://127.0.0.1:8080/health"]
+      interval: 5s
+      timeout: 3s
+      retries: 40
+      start_period: 10s
+
+  image-prompt-enhancer-t2i:
+    image: mike-ai/llama.cpp:b10930
+    container_name: mike-ai-image-prompt-enhancer-t2i
+    restart: "no"
+    profiles: [image]
+    labels:
+      com.mike-ai.image-worker: image-prompt-t2i
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ["${QWEN_IMAGE_PE_GPU:-0}"]
+              capabilities: [gpu]
+    read_only: true
+    tmpfs: ["/tmp:size=512m,mode=1777"]
+    volumes:
+      - "${QWEN_IMAGE_PE_T2I_MODEL_DIR:-/data/models/qwen-image-2.1-pe-t2i-q5}:/models:ro"
+    command:
+      - --model
+      - /models/Qwen-Image-2.1-PE-T2I.Q5_K_M.gguf
+      - --alias
+      - qwen-image-pe-t2i
+      - --ctx-size
+      - "16384"
+      - --flash-attn
+      - "on"
+      - --cache-type-k
+      - q8_0
+      - --cache-type-v
+      - q8_0
+      - --threads
+      - "6"
+      - --threads-batch
+      - "6"
+      - --batch-size
+      - "1024"
+      - --ubatch-size
+      - "128"
+      - --parallel
+      - "1"
+      - --jinja
+      - --reasoning
+      - auto
+      - --host
+      - 0.0.0.0
+      - --port
+      - "8080"
+      - --metrics
+      - --n-gpu-layers
+      - all
+      - --device
+      - CUDA0
+      - --split-mode
+      - none
+      - --no-ui
+    networks: [inference]
+    security_opt: ["no-new-privileges:true"]
+    cap_drop: [ALL]
+    healthcheck:
+      test: [CMD, curl, -fsS, "http://127.0.0.1:8080/health"]
+      interval: 5s
+      timeout: 3s
+      retries: 40
+      start_period: 10s
+
   # Previous production image model, retained as a stopped rollback target.
   # It is outside the normal router path and can only be started through the
   # controller's allowlisted flux-standby endpoint.

dev/test_router_coordination.py

@@ -13,6 +13,18 @@ import ai_profile_router as router
 
 
 class CoordinationTests(unittest.TestCase):
+    def test_prompt_enhancer_accepts_plain_and_fenced_json(self):
+        plain = router._parse_prompt_enhancer_result(
+            '{"rewritten_prompt":"new scene","wh_ratio":"3:2"}')
+        fenced = router._parse_prompt_enhancer_result(
+            '```json\n{"rewritten_prompt":"portrait","wh_ratio":"3:4"}\n```')
+        self.assertEqual(plain['rewritten_prompt'], 'new scene')
+        self.assertEqual(fenced['wh_ratio'], '3:4')
+
+    def test_prompt_enhancer_rejects_missing_rewrite(self):
+        with self.assertRaisesRegex(RuntimeError, 'rewritten_prompt'):
+            router._parse_prompt_enhancer_result('{"wh_ratio":"1:1"}')
+
     def test_mode_uses_one_consistent_controller_snapshot(self):
         with patch.object(router, 'PROFILE_CONTROL_URL', 'http://controller'), \
              patch.object(router, '_profile_controller_request', return_value={

platform/docker/profile-controller/profile_controller.py

@@ -27,6 +27,10 @@ LABEL_KEY = "com.mike-ai.llama-profile"
 IMAGE_LABEL_KEY = "com.mike-ai.image-worker"
 IMAGE_WORKER = os.environ.get("IMAGE_WORKER", "image")
 RESTORE_WORKER = os.environ.get("RESTORE_WORKER", "restore")
+IMAGE_PROMPT_I2I_WORKER = os.environ.get(
+    "IMAGE_PROMPT_I2I_WORKER", "image-prompt-i2i").strip()
+IMAGE_PROMPT_T2I_WORKER = os.environ.get(
+    "IMAGE_PROMPT_T2I_WORKER", "image-prompt-t2i").strip()
 FLUX_STANDBY_WORKER = os.environ.get(
     "FLUX_STANDBY_WORKER", "flux-standby").strip()
 TTS_LABEL_KEY = "com.mike-ai.tts-worker"
@@ -102,7 +106,12 @@ def image_container(kind: str = IMAGE_WORKER) -> dict:
 
 def image_containers() -> list[dict]:
     """All allowlisted GPU workers that must never overlap an LLM."""
-    allowed = {IMAGE_WORKER, RESTORE_WORKER}
+    allowed = {
+        IMAGE_WORKER,
+        RESTORE_WORKER,
+        IMAGE_PROMPT_I2I_WORKER,
+        IMAGE_PROMPT_T2I_WORKER,
+    }
     if FLUX_STANDBY_WORKER:
         allowed.add(FLUX_STANDBY_WORKER)
     return [item for item in labelled_containers(IMAGE_LABEL_KEY)
@@ -363,7 +372,13 @@ def stop_inference() -> dict:
 
 
 def set_image_worker(running: bool, kind: str = IMAGE_WORKER) -> dict:
-    if kind not in {IMAGE_WORKER, RESTORE_WORKER, FLUX_STANDBY_WORKER}:
+    if kind not in {
+        IMAGE_WORKER,
+        RESTORE_WORKER,
+        FLUX_STANDBY_WORKER,
+        IMAGE_PROMPT_I2I_WORKER,
+        IMAGE_PROMPT_T2I_WORKER,
+    }:
         raise ValueError("worker is not allowlisted")
     with LOCK:
         item = image_container(kind)
@@ -799,6 +814,10 @@ class Handler(BaseHTTPRequestHandler):
         worker_paths = {
             "/workers/image/start": (IMAGE_WORKER, True),
             "/workers/image/stop": (IMAGE_WORKER, False),
+            "/workers/image-prompt-i2i/start": (IMAGE_PROMPT_I2I_WORKER, True),
+            "/workers/image-prompt-i2i/stop": (IMAGE_PROMPT_I2I_WORKER, False),
+            "/workers/image-prompt-t2i/start": (IMAGE_PROMPT_T2I_WORKER, True),
+            "/workers/image-prompt-t2i/stop": (IMAGE_PROMPT_T2I_WORKER, False),
             "/workers/restore/start": (RESTORE_WORKER, True),
             "/workers/restore/stop": (RESTORE_WORKER, False),
             "/workers/flux-standby/start": (FLUX_STANDBY_WORKER, True),

router/ai_profile_router.py

@@ -146,6 +146,18 @@ 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_PROMPT_I2I_URL = os.environ.get(
+    "IMAGE_PROMPT_I2I_URL", "").rstrip("/")
+IMAGE_PROMPT_T2I_URL = os.environ.get(
+    "IMAGE_PROMPT_T2I_URL", "").rstrip("/")
+IMAGE_PROMPT_I2I_SYSTEM_FILE = os.environ.get(
+    "IMAGE_PROMPT_I2I_SYSTEM_FILE",
+    "/etc/mike-ai/qwen-image-pe-i2i-system-prompt.txt")
+IMAGE_PROMPT_T2I_SYSTEM_FILE = os.environ.get(
+    "IMAGE_PROMPT_T2I_SYSTEM_FILE",
+    "/etc/mike-ai/qwen-image-pe-t2i-system-prompt.txt")
+IMAGE_PROMPT_ENHANCER_TIMEOUT = float(os.environ.get(
+    "IMAGE_PROMPT_ENHANCER_TIMEOUT", "180"))
 IMAGE_MODEL_NAME = os.environ.get(
     "IMAGE_MODEL_NAME", "Qwen-Image-2.1-int8")
 IMAGE_INFERENCE_STEPS = int(os.environ.get("IMAGE_INFERENCE_STEPS", "25"))
@@ -182,6 +194,16 @@ IMAGE_QUALITY = {"standard": IMAGE_INFERENCE_STEPS,
 IMAGE_DEFAULT_QUALITY = "standard"
 IMAGE_MAX_N = 4
 
+IMAGE_RATIO_SIZES = {
+    "1:1": (1024, 1024),
+    "3:2": (1536, 1024),
+    "2:3": (1024, 1536),
+    "4:3": (1536, 1024),
+    "3:4": (1024, 1536),
+    "16:9": (1920, 1088),
+    "9:16": (1088, 1920),
+}
+
 # --- Sprachausgabe (Qwen3-TTS über das interne Normalisierungs-Gateway) ---
 TTS_WORKER_URL = os.environ.get("TTS_WORKER_URL", "http://127.0.0.1:8085")
 TTS_TIMEOUT = float(os.environ.get("TTS_TIMEOUT", "300"))  # s, pro Synthese
@@ -1386,11 +1408,146 @@ def _restore_qwen(profile: str) -> None:
     RUNTIME.save(last_profile=profile, phase="idle")
 
 
+def _image_data_url(path: str) -> str:
+    """Read one already validated local reference image as a data URL."""
+    with open(path, "rb") as handle:
+        data = handle.read()
+    if data.startswith(b"\x89PNG\r\n\x1a\n"):
+        mime = "image/png"
+    elif data.startswith(b"\xff\xd8\xff"):
+        mime = "image/jpeg"
+    elif data.startswith(b"RIFF") and data[8:12] == b"WEBP":
+        mime = "image/webp"
+    else:
+        raise RuntimeError(f"Referenzbild hat ein unbekanntes Format: {path}")
+    return f"data:{mime};base64,{base64.b64encode(data).decode()}"
+
+
+def _parse_prompt_enhancer_result(content: object) -> dict:
+    if not isinstance(content, str) or not content.strip():
+        raise RuntimeError("Prompt-Enhancer lieferte keine Antwort")
+    text = content.strip()
+    if text.startswith("```"):
+        text = re.sub(r"^```(?:json)?\s*", "", text, flags=re.IGNORECASE)
+        text = re.sub(r"\s*```$", "", text)
+    try:
+        result = json.loads(text)
+    except ValueError:
+        start, end = text.find("{"), text.rfind("}")
+        if start < 0 or end <= start:
+            raise RuntimeError("Prompt-Enhancer lieferte kein JSON")
+        try:
+            result = json.loads(text[start:end + 1])
+        except ValueError as exc:
+            raise RuntimeError("Prompt-Enhancer lieferte ungültiges JSON") from exc
+    if not isinstance(result, dict):
+        raise RuntimeError("Prompt-Enhancer lieferte kein JSON-Objekt")
+    rewritten = result.get("rewritten_prompt")
+    if not isinstance(rewritten, str) or not rewritten.strip():
+        raise RuntimeError("Prompt-Enhancer lieferte keinen rewritten_prompt")
+    if len(rewritten) > 8000:
+        raise RuntimeError("Aufbereiteter Bildprompt ist länger als 8000 Zeichen")
+    result["rewritten_prompt"] = rewritten.strip()
+    return result
+
+
+def _enhance_image_prompt(prompt: str, source_files: list[str]) -> tuple[str, dict]:
+    """Run the official Qwen Image 2.1 prompt enhancer on the RTX 3060."""
+    editing = bool(source_files)
+    kind = "image-prompt-i2i" if editing else "image-prompt-t2i"
+    url = IMAGE_PROMPT_I2I_URL if editing else IMAGE_PROMPT_T2I_URL
+    system_file = (IMAGE_PROMPT_I2I_SYSTEM_FILE if editing
+                   else IMAGE_PROMPT_T2I_SYSTEM_FILE)
+    model = "qwen-image-pe-i2i" if editing else "qwen-image-pe-t2i"
+    if not PROFILE_CONTROL_URL or not url:
+        raise RuntimeError("Qwen-Image-Prompt-Enhancer ist nicht konfiguriert")
+    try:
+        with open(system_file, encoding="utf-8") as handle:
+            system_prompt = handle.read().strip()
+    except OSError as exc:
+        raise RuntimeError(f"Systemprompt des Prompt-Enhancers fehlt: {exc}") from exc
+
+    started = time.monotonic()
+    _profile_controller_request("POST", f"/workers/{kind}/start")
+    try:
+        deadline = time.monotonic() + IMAGE_START_TIMEOUT
+        while True:
+            try:
+                with urllib.request.urlopen(url + "/health", timeout=3) as response:
+                    health = json.load(response)
+                if health.get("status") == "ok":
+                    break
+            except (OSError, urllib.error.URLError, TimeoutError, ValueError):
+                pass
+            if time.monotonic() >= deadline:
+                raise RuntimeError("Prompt-Enhancer hat nicht gestartet")
+            time.sleep(1)
+
+        if editing:
+            user_content: object = [
+                {"type": "image_url", "image_url": {"url": _image_data_url(path)}}
+                for path in source_files
+            ]
+            user_content.append({"type": "text", "text": prompt})
+        else:
+            user_content = prompt
+        request_body = {
+            "model": model,
+            "messages": [
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": user_content},
+            ],
+            "temperature": 1.0,
+            "top_p": 0.95,
+            "top_k": 20,
+            "presence_penalty": 0.0 if editing else 1.5,
+            "max_tokens": 4096,
+            "thinking_budget_tokens": 2048,
+            "chat_template_kwargs": {
+                "enable_thinking": True,
+                "reasoning_effort": "low",
+            },
+        }
+        req = urllib.request.Request(
+            url + "/v1/chat/completions",
+            data=json.dumps(request_body).encode(),
+            method="POST",
+            headers={"Content-Type": "application/json"},
+        )
+        try:
+            with urllib.request.urlopen(
+                    req, timeout=IMAGE_PROMPT_ENHANCER_TIMEOUT) as response:
+                payload = json.load(response)
+        except urllib.error.HTTPError as exc:
+            detail = exc.read(4096).decode(errors="replace")
+            raise RuntimeError(
+                f"Prompt-Enhancer HTTP {exc.code}: {detail[-500:]}") from exc
+        try:
+            content = payload["choices"][0]["message"]["content"]
+        except (KeyError, IndexError, TypeError) as exc:
+            raise RuntimeError("Prompt-Enhancer-Antwort ist unvollständig") from exc
+        result = _parse_prompt_enhancer_result(content)
+        metadata = {
+            "model": model,
+            "quantization": "Q5_K_M",
+            "seconds": round(time.monotonic() - started, 3),
+            "wh_ratio": result.get("wh_ratio"),
+            "ratio_follow": result.get("ratio_follow"),
+        }
+        return result["rewritten_prompt"], metadata
+    finally:
+        try:
+            _profile_controller_request("POST", f"/workers/{kind}/stop")
+        except Exception as exc:
+            log.warning("Prompt-Enhancer ließ sich nicht stoppen: %s", exc)
+
+
 def generate_image(prompt: str, width: int, height: int, steps: int,
                    guidance: float, seed: int | None, n: int,
                    quality: str = "standard",
                    source_files: list[str] | None = None,
                    model: str = IMAGE_MODEL_NAME,
+                   size_explicit: bool = False,
                    ) -> tuple[list[str], str | None]:
     """Orchestriert die Bildgenerierung inkl. Qwen-Hotswap.
 
@@ -1411,6 +1568,8 @@ def generate_image(prompt: str, width: int, height: int, steps: int,
         warning: str | None = None
         img.last_error = None
         img.current_model = model
+        original_prompt = prompt
+        prompt_enhancer: dict | None = None
         # Qwen wird gestoppt → für Chats nicht verfügbar (die warten).
         _set_qwen_unavailable(True)
         try:
@@ -1433,11 +1592,26 @@ def generate_image(prompt: str, width: int, height: int, steps: int,
                         f"(Exit {proc.returncode}): {out[-500:]}")
             _wait_upstream_down(time.monotonic() + 60)
 
-            # 2) Worker starten (Modell wird beim ersten generate geladen).
+            # 2) Den knappen Benutzerprompt mit dem offiziellen Qwen-
+            # Prompt-Enhancer auf der RTX 3060 in einen belastbaren
+            # Produktionsprompt umschreiben. Danach wird der Enhancer wieder
+            # beendet, bevor der eigentliche Bildworker startet.
+            img.phase = "enhancing-prompt"
+            prompt, prompt_enhancer = _enhance_image_prompt(
+                original_prompt, source_files or [])
+            ratio = prompt_enhancer.get("wh_ratio")
+            if not size_explicit and isinstance(ratio, str):
+                width, height = IMAGE_RATIO_SIZES.get(
+                    ratio.strip(), (width, height))
+            log.info("Bildprompt aufbereitet (%s, %.1f s, Verhältnis %s)",
+                     prompt_enhancer["model"],
+                     prompt_enhancer["seconds"], ratio)
+
+            # 3) Worker starten (Modell wird beim ersten generate geladen).
             img.phase = "loading-image"
             worker = _worker()
 
-            # 3) Generieren.
+            # 4) Generieren.
             for i in range(n):
                 img.phase = "generating"
                 filename = time.strftime("%Y%m%d-%H%M%S") + \
@@ -1465,6 +1639,8 @@ def generate_image(prompt: str, width: int, height: int, steps: int,
                 # Metadaten speichern (Sidecar-JSON).
                 meta = {
                     "prompt": prompt,
+                    "original_prompt": original_prompt,
+                    "prompt_enhancer": prompt_enhancer,
                     "seed": seed,
                     "width": width,
                     "height": height,
@@ -1493,7 +1669,7 @@ def generate_image(prompt: str, width: int, height: int, steps: int,
             if removed:
                 log.info("Bild-Retention: %d alte Bilder entfernt", len(removed))
 
-            # 4) Worker vollständig beenden (VRAM + CUDA-Kontext freigeben).
+            # 5) Worker vollständig beenden (VRAM + CUDA-Kontext freigeben).
             img.phase = "unloading-image"
             worker.stop()
             img.worker = None
@@ -1510,7 +1686,7 @@ def generate_image(prompt: str, width: int, height: int, steps: int,
                 img.worker = None
             raise
         finally:
-            # 5) Qwen immer wiederherstellen.
+            # 6) Qwen immer wiederherstellen.
             img.phase = "restoring-qwen"
             try:
                 _restore_qwen(profile)
@@ -2358,6 +2534,7 @@ class Handler(BaseHTTPRequestHandler):
             return
 
         # Größe
+        size_explicit = "size" in data
         size = data.get("size", "1024x1024")
         if size not in IMAGE_SIZES:
             self._send_error(
@@ -2425,7 +2602,7 @@ class Handler(BaseHTTPRequestHandler):
         try:
             results, warning = generate_image(
                 prompt.strip(), width, height, steps, guidance, seed, n,
-                quality, source_files, model)
+                quality, source_files, model, size_explicit)
         except (ValueError, RuntimeError) as e:
             self._send_error(503, str(e), "server_error", "image_generation_failed")
             return

scripts/prepare-qwen-image-21.sh

@@ -5,6 +5,8 @@ set -Eeuo pipefail
 ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 ENV_FILE=${STACK_ENV:-/etc/mike-ai/stack.env}
 MODEL_DIR=${QWEN_IMAGE_21_MODEL_DIR:-/data/models/Qwen-Image-2.1-ComfyUI}
+PE_I2I_DIR=${QWEN_IMAGE_PE_I2I_MODEL_DIR:-/data/models/qwen-image-2.1-pe-i2i-q5}
+PE_T2I_DIR=${QWEN_IMAGE_PE_T2I_MODEL_DIR:-/data/models/qwen-image-2.1-pe-t2i-q5}
 BASE=https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/ace0edeb3791a594ddfa36ed5f41a178a394e921
 
 download() {
@@ -19,15 +21,52 @@ download() {
   mv "$target.part" "$target"
 }
 
+download_url() {
+  local url=$1 target=$2 expected=$3
+  install -d -m 0755 "$(dirname "$target")"
+  if [[ -f $target ]] && echo "$expected  $target" | sha256sum -c --status; then
+    echo "OK: $target"
+    return
+  fi
+  curl -fL --retry 5 --retry-all-errors -C - -o "$target.part" "$url"
+  echo "$expected  $target.part" | sha256sum -c --status
+  mv "$target.part" "$target"
+}
+
 download diffusion_models/qwen_image_2.1_int8_convrot.safetensors cb74113cb03faecd79611b01fd7fd642f0aa60d6f0b95086abee214d75eaa57d
 download text_encoders/qwen3vl_8b_int8_convrot.safetensors 8bfd0f6e12abf2d2d697ecc888e5e90b0d6741d6708f05799f53afa560452e8f
 download vae/qwen_image_2.1_vae_bf16.safetensors bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9
 
+# Official Qwen Image 2.1 Prompt Enhancers, quantized to Q5_K_M for the
+# 12-GiB RTX 3060. I2I uses the separate multimodal projector; T2I does not.
+download_url \
+  https://huggingface.co/prithivMLmods/Qwen-Image-2.1-PE-I2I-GGUF/resolve/55b9c1a326599e142d59bcad8715d5601ccf8daa/Qwen-Image-2.1-PE-I2I.Q5_K_M.gguf \
+  "$PE_I2I_DIR/Qwen-Image-2.1-PE-I2I.Q5_K_M.gguf" \
+  cb71f71fe5fe5938570d65a7c403fbcb1a9065dff9dd9a021f58aec42785b84e
+download_url \
+  https://huggingface.co/prithivMLmods/Qwen-Image-2.1-PE-I2I-GGUF/resolve/55b9c1a326599e142d59bcad8715d5601ccf8daa/Qwen-Image-2.1-PE-I2I.mmproj-bf16.gguf \
+  "$PE_I2I_DIR/Qwen-Image-2.1-PE-I2I.mmproj-bf16.gguf" \
+  8dedb71dbc3092dc47de9108ad373d68a12854e2527d59bd9399601738f3bce1
+download_url \
+  https://huggingface.co/Qwen/Qwen-Image-2.1-PE-I2I/resolve/72927bc08afc99b7888ceb7d7d51a12db3700bbd/system_prompt.txt \
+  "$PE_I2I_DIR/system_prompt.txt" \
+  e378fea686a1431581ba4c654d332ae96adad633f144ae738ec8ce9c4fd66439
+download_url \
+  https://huggingface.co/prithivMLmods/Qwen-Image-2.1-PE-T2I-GGUF/resolve/e18d4a3e0830ab157770738b16830e6fcf5f57d4/Qwen-Image-2.1-PE-T2I.Q5_K_M.gguf \
+  "$PE_T2I_DIR/Qwen-Image-2.1-PE-T2I.Q5_K_M.gguf" \
+  749f5652fd6e8b760ca860091f7a5bffa254a7954203ae5c9bcdf5f93197702f
+download_url \
+  https://huggingface.co/Qwen/Qwen-Image-2.1-PE-T2I/resolve/f3ed7985c788ad75b3ab7223e0c4c51e2a43545b/system_prompt.txt \
+  "$PE_T2I_DIR/system_prompt.txt" \
+  a77c9a06c59b120741141d9514b95682bb8761d02bec49ca61def7b2b3d9fb99
+
 compose=(docker compose --env-file "$ENV_FILE" -f "$ROOT_DIR/compose.yaml" --profile image --profile flux-standby)
 "${compose[@]}" build image-worker
 "${compose[@]}" up -d --build --no-deps profile-controller
-"${compose[@]}" create image-worker flux-image-worker
-"${compose[@]}" stop image-worker flux-image-worker
+"${compose[@]}" create image-worker image-prompt-enhancer-i2i \
+  image-prompt-enhancer-t2i flux-image-worker
+"${compose[@]}" stop image-worker image-prompt-enhancer-i2i \
+  image-prompt-enhancer-t2i flux-image-worker
 state=$(docker inspect -f '{{.State.Status}}' mike-ai-image-worker)
 [[ $state == exited || $state == created ]]
-echo "Qwen-Image-2.1 production worker and FLUX standby are prepared and stopped."
+echo "Qwen-Image-2.1, both prompt enhancers and the FLUX standby are prepared and stopped."