diff --git a/backend/ai_server.py b/backend/ai_server.py
index fdd71d7..5dc135d 100644
--- a/backend/ai_server.py
+++ b/backend/ai_server.py
@@ -1607,12 +1607,21 @@ def predict_batch(req: PredictBatchRequest):
             "error": "no paths supplied",
         }
 
+    imgsz = int(req.imageSize or _IMGSZ or 640)
+
     current_model = get_model()
     if current_model is None:
+        predictions = [empty_prediction("detector_model_missing") for _ in paths]
+        if not req.detectorOnly:
+            apply_pose_batch_to_predictions(paths, predictions, imgsz)
+
         return {
             "ok": True,
-            "predictions": [empty_prediction("model_missing") for _ in paths],
-            "error": _MODEL_ERROR or f"YOLO model not found: {DEFAULT_MODEL_PATH}",
+            "available": False,
+            "modelAvailable": False,
+            "predictions": predictions,
+            "modelError": _MODEL_ERROR,
+            "expectedModel": str(DEFAULT_MODEL_PATH),
         }
 
     if DETECTION_LABELS_PATH is None or _LABEL_ERROR:
@@ -1622,8 +1631,6 @@ def predict_batch(req: PredictBatchRequest):
             "error": f"detection labels missing: {_LABEL_ERROR}",
         }
 
-    imgsz = int(req.imageSize or _IMGSZ or 640)
-
     try:
         results = current_model.predict(
             source=paths,
@@ -1642,6 +1649,8 @@ def predict_batch(req: PredictBatchRequest):
 
         return {
             "ok": True,
+            "available": True,
+            "modelAvailable": True,
             "predictions": predictions,
         }
 
@@ -1703,7 +1712,7 @@ def health():
 
     status_payload = {
         "ok": True,
-        "ready": current_model is not None,
+        "ready": True,
         "modelAvailable": current_model is not None,
         "model": _MODEL_PATH,
         "modelError": _MODEL_ERROR,
@@ -1751,8 +1760,8 @@ def reload_model():
     names = getattr(current_model, "names", {}) or {} if current_model is not None else {}
 
     status_payload = {
-        "ok": current_model is not None,
-        "ready": current_model is not None,
+        "ok": True,
+        "ready": True,
         "modelAvailable": current_model is not None,
         "model": _MODEL_PATH,
         "modelError": _MODEL_ERROR,
diff --git a/backend/dist/nsfwapp-linux-amd64 b/backend/dist/nsfwapp-linux-amd64
index 15400ac..3b43f25 100644
Binary files a/backend/dist/nsfwapp-linux-amd64 and b/backend/dist/nsfwapp-linux-amd64 differ
diff --git a/backend/dist/nsfwapp.exe b/backend/dist/nsfwapp.exe
index b180390..010066f 100644
Binary files a/backend/dist/nsfwapp.exe and b/backend/dist/nsfwapp.exe differ
diff --git a/backend/dist/nsfwapp_amd64.deb b/backend/dist/nsfwapp_amd64.deb
index bc353f7..dbe5cee 100644
Binary files a/backend/dist/nsfwapp_amd64.deb and b/backend/dist/nsfwapp_amd64.deb differ
diff --git a/backend/ml_setup.go b/backend/ml_setup.go
index 93cd651..7c422c1 100644
--- a/backend/ml_setup.go
+++ b/backend/ml_setup.go
@@ -9,6 +9,7 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"runtime"
 	"strings"
 	"sync"
 	"time"
@@ -112,11 +113,22 @@ func mlPythonSetupConfigFromEnv() mlPythonSetupConfig {
 		appDir:           filepath.Clean(appDir),
 		requirementsPath: filepath.Clean(requirementsPath),
 		venvDir:          filepath.Clean(venvDir),
-		venvPython:       filepath.Join(filepath.Clean(venvDir), "bin", "python"),
+		venvPython:       mlPythonVenvPythonPath(venvDir),
 		requirementsMark: filepath.Join(filepath.Clean(venvDir), ".requirements-ml.txt"),
 	}
 }
 
+func mlPythonVenvPythonPath(venvDir string) string {
+	venvDir = filepath.Clean(strings.TrimSpace(venvDir))
+	if venvDir == "" {
+		return ""
+	}
+	if runtime.GOOS == "windows" {
+		return filepath.Join(venvDir, "Scripts", "python.exe")
+	}
+	return filepath.Join(venvDir, "bin", "python")
+}
+
 func defaultMLPythonVenvDir() string {
 	base := strings.TrimSpace(os.TempDir())
 	if base == "" {
diff --git a/backend/server.go b/backend/server.go
index a1c02ae..2585f9b 100644
--- a/backend/server.go
+++ b/backend/server.go
@@ -206,6 +206,10 @@ func aiServerPythonPath() string {
 		return raw
 	}
 
+	if venvPython := configuredMLPythonVenvPythonPath(); venvPython != "" {
+		return venvPython
+	}
+
 	// Windows py launcher zuerst versuchen.
 	if runtime.GOOS == "windows" {
 		if _, err := exec.LookPath("py"); err == nil {
@@ -224,6 +228,34 @@ func aiServerPythonPath() string {
 	return "python"
 }
 
+func configuredMLPythonVenvPythonPath() string {
+	if strings.TrimSpace(os.Getenv("NSFWAPP_SKIP_ML_SETUP")) == "1" {
+		return ""
+	}
+
+	candidates := []string{}
+	if venvDir := strings.TrimSpace(os.Getenv("NSFWAPP_ML_VENV")); venvDir != "" {
+		candidates = append(candidates, venvDir)
+	}
+	if strings.TrimSpace(os.Getenv("NSFWAPP_ML_SETUP")) == "1" {
+		candidates = append(candidates, defaultMLPythonVenvDir())
+	}
+
+	seen := map[string]bool{}
+	for _, dir := range candidates {
+		path := mlPythonVenvPythonPath(dir)
+		key := filepath.Clean(path)
+		if path == "" || seen[key] {
+			continue
+		}
+		seen[key] = true
+		if info, err := os.Stat(path); err == nil && info != nil && !info.IsDir() {
+			return path
+		}
+	}
+	return ""
+}
+
 func findAIServerScriptDir() (string, error) {
 	cwd, _ := os.Getwd()
 
@@ -544,7 +576,7 @@ func startAIServer(ctx context.Context) (*aiServerProcess, error) {
 
 	// Modell ebenfalls standardmäßig aus generated/training nehmen,
 	// aber YOLO_MODEL darf weiterhin extern überschrieben werden.
-	if strings.TrimSpace(os.Getenv("YOLO_MODEL")) == "" {
+	if strings.TrimSpace(os.Getenv("YOLO_MODEL")) == "" && defaultModel.EffectiveExists {
 		env = upsertEnv(env, "YOLO_MODEL", defaultModelPath)
 	}
 
diff --git a/backend/training.go b/backend/training.go
index c77b792..779e1c3 100644
--- a/backend/training.go
+++ b/backend/training.go
@@ -2557,7 +2557,38 @@ func trainingCancelHandler(w http.ResponseWriter, r *http.Request) {
 }
 
 func trainingRunJob(ctx context.Context, root string, count int) {
+	if err := ensureMLPythonSetup(ctx); err != nil {
+		if errors.Is(err, context.Canceled) {
+			trainingFinishCancelled(root)
+			return
+		}
+
+		appLogln("ML-Python setup for training failed:", err)
+		trainingSetJobStatus(func(s *TrainingJobStatus) {
+			finishedAt := time.Now().UTC()
+			var durationMs int64
+			if startedAt, parseErr := time.Parse(time.RFC3339, strings.TrimSpace(s.StartedAt)); parseErr == nil {
+				durationMs = finishedAt.Sub(startedAt).Milliseconds()
+				if durationMs < 0 {
+					durationMs = 0
+				}
+			}
+
+			s.Running = false
+			s.Progress = 100
+			s.Step = "Training fehlgeschlagen."
+			s.Message = "ML-Python-Umgebung konnte nicht vorbereitet werden."
+			s.Error = err.Error()
+			s.FinishedAt = finishedAt.Format(time.RFC3339)
+			s.DurationMs = durationMs
+			s.PreviewURL = ""
+		})
+		trainingClearJobCancel()
+		return
+	}
+
 	python := trainingPythonExe()
+	appLogln("ML-Python für Training:", python)
 
 	cleanOutput := func(text string) string {
 		out := strings.TrimSpace(text)
@@ -6504,7 +6535,7 @@ func trainingPythonExe() string {
 	if v != "" {
 		return v
 	}
-	return "python"
+	return aiServerPythonPath()
 }
 
 func trainingProjectRoot() string {