diff --git a/modules/trainer/BaseTrainer.py b/modules/trainer/BaseTrainer.py
index 20fc5eb7a..e4b7a3b29 100644
--- a/modules/trainer/BaseTrainer.py
+++ b/modules/trainer/BaseTrainer.py
@@ -97,7 +97,13 @@ def _start_tensorboard(self):
         if self.config.tensorboard_expose:
             tensorboard_args.append("--bind_all")
 
-        self.tensorboard_subprocess = subprocess.Popen(tensorboard_args)
+        # Discard the tensorboard child's stdout/stderr: the TF-not-found notice, the
+        # experimental-data-loading NOTE and the serving banner are all noise, and the
+        # UI already exposes the tensorboard URL. Popen still raises if the executable
+        # is missing, so a real launch failure is not hidden.
+        self.tensorboard_subprocess = subprocess.Popen(
+            tensorboard_args, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
+        )
 
     def _stop_tensorboard(self):
         self.tensorboard_subprocess.kill()
diff --git a/modules/ui/TrainUIController.py b/modules/ui/TrainUIController.py
index 72623def4..2c4896226 100644
--- a/modules/ui/TrainUIController.py
+++ b/modules/ui/TrainUIController.py
@@ -103,8 +103,13 @@ def _start_always_on_tensorboard(self):
         if self.train_config.tensorboard_expose:
             tensorboard_args.append("--bind_all")
 
+        # Discard the tensorboard child's stdout/stderr: the TF-not-found notice, the
+        # experimental-data-loading NOTE and the serving banner are all noise, and the
+        # UI already exposes the tensorboard URL.
         try:
-            self.always_on_tensorboard_subprocess = subprocess.Popen(tensorboard_args)
+            self.always_on_tensorboard_subprocess = subprocess.Popen(
+                tensorboard_args, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
+            )
         except Exception:
             self.always_on_tensorboard_subprocess = None
 
diff --git a/modules/util/ui/pyside6_util.py b/modules/util/ui/pyside6_util.py
index fd7b2ef8a..42b4a3e60 100644
--- a/modules/util/ui/pyside6_util.py
+++ b/modules/util/ui/pyside6_util.py
@@ -1,3 +1,4 @@
+import os
 import signal
 import sys
 from abc import ABCMeta
@@ -19,6 +20,14 @@ def create_application() -> QApplication:
     # active and Ctrl+C would be ignored.
     signal.signal(signal.SIGINT, signal.SIG_DFL)
 
+    # On desktops without the xdg-desktop-portal Settings interface, Qt spams two
+    # "qt.qpa.theme.gnome: dbus reply error ... org.freedesktop.portal.Settings"
+    # lines while probing for the system color scheme. Silence just that category;
+    # the rules string is read when Qt's logging initializes at QApplication init.
+    _gnome_theme_rule = "qt.qpa.theme.gnome=false"
+    existing_rules = os.environ.get("QT_LOGGING_RULES")
+    os.environ["QT_LOGGING_RULES"] = f"{existing_rules};{_gnome_theme_rule}" if existing_rules else _gnome_theme_rule
+
     app = QApplication(sys.argv)
     # Force Fusion everywhere: native styles (e.g. windowsvista) draw standard
     # controls via OS theme APIs, which breaks once an application stylesheet
diff --git a/scripts/util/import_util.py b/scripts/util/import_util.py
index 150df4d8a..a0c0a83b6 100644
--- a/scripts/util/import_util.py
+++ b/scripts/util/import_util.py
@@ -1,7 +1,9 @@
 def script_imports(allow_zluda: bool = True):
     import logging
     import os
+    import re
     import sys
+    import warnings
     from pathlib import Path
 
     # Filter out the Triton warning on startup.
@@ -10,6 +12,40 @@ def script_imports(allow_zluda: bool = True):
         .getLogger("xformers") \
         .addFilter(lambda record: 'A matching Triton is not available' not in record.getMessage())
 
+    # Silence specific non-actionable startup/compile warnings. A logger filter
+    # targets the exact emitting logger, since a parent logger's filter misses
+    # records from child loggers.
+
+    # diffusers/transformers chatty logger.warning() lines at import/load time.
+    logging.getLogger("diffusers.modular_pipelines").addFilter(
+        lambda record: 'Modular Diffusers is currently an experimental feature' not in record.getMessage()
+    )
+    # The subject of these two is interpolated into the message, so match the whole
+    # sentence with .* standing in for the runtime value.
+    logging.getLogger("diffusers.configuration_utils").addFilter(
+        lambda record: not re.search(
+            r"The config attributes .* were passed to .*, but are not expected and will be ignored",
+            record.getMessage(),
+        )
+    )
+    logging.getLogger("transformers.modeling_utils").addFilter(
+        lambda record: not re.search(
+            r"`loss_type=.*` was set in the config but it is unrecognized", record.getMessage()
+        )
+    )
+
+    # A dependency still calls hf_hub_download with the removed local_dir_use_symlinks
+    # argument; the deprecation warning is not actionable.
+    warnings.filterwarnings("ignore", message=r".*local_dir_use_symlinks.*")
+
+    # torch.compile emits performance notes when inductor falls back or can't use a
+    # fast path; harmless and noisy for normal runs. The SMs note is a logger.warning()
+    # on its exact emitting logger; the complex-operators note is a warnings.warn().
+    warnings.filterwarnings("ignore", message=r".*does not support code generation for complex operators.*")
+    logging.getLogger("torch._inductor.utils").addFilter(
+        lambda record: 'Not enough SMs to use max_autotune_gemm mode' not in record.getMessage()
+    )
+
     # Insert ourselves as the highest-priority library path, so our modules are
     # always found without any risk of being shadowed by another import path.
     # 3 .parent calls to navigate from /scripts/util/import_util.py to the main directory