diff --git a/joss/paper.pdf b/joss/paper.pdf
index ee8e792..df425e8 100644
Binary files a/joss/paper.pdf and b/joss/paper.pdf differ
diff --git a/tfaip/predict/predictorbase.py b/tfaip/predict/predictorbase.py
index afe0694..f07893c 100644
--- a/tfaip/predict/predictorbase.py
+++ b/tfaip/predict/predictorbase.py
@@ -30,6 +30,7 @@
 from tfaip.data.pipeline.datagenerator import DataGenerator
 from tfaip.data.pipeline.datapipeline import DataPipeline
 from tfaip.device.device_config import DeviceConfig, distribute_strategy
+from tfaip.predict.raw_predictor import RawPredictor
 from tfaip.trainer.callbacks.benchmark_callback import BenchmarkResults
 from tfaip.util.multiprocessing.parallelmap import tqdm_wrapper
 from tfaip.util.profiling import MeasureTime
@@ -95,6 +96,27 @@ def _load_model(self, model: Union[str, keras.Model]):
 
         return model
 
+    def raw(self) -> RawPredictor:
+        """Create a raw predictor from this predictor that allows to asynchronly predict raw samples.
+
+        Usage:
+
+        Either call
+
+        with predictor.raw() as raw_pred:
+            raw_pred(sample1)
+            raw_pred(sample2)
+            ...
+
+        or
+
+        raw_pred = predictor.raw().__enter__()
+        raw_pred(sample1)
+        raw_pred(sample2)
+
+        """
+        return RawPredictor(self)
+
     def predict(self, params: DataGeneratorParams) -> Iterable[Sample]:
         """
         Predict a DataGenerator based on its params.
@@ -134,6 +156,7 @@ def predict_raw(self, inputs: Iterable[Any], *, size=None) -> Iterable[Sample]:
             - predict_pipeline
             - predict_dataset
             - predict
+            - raw
         """
         if size is None:
             # Automatically compute the size (number of samples)
@@ -159,10 +182,11 @@ def create_data_generator(self) -> DataGenerator:
                 return RawGenerator(mode=self.mode, params=self.generator_params)
 
         pipeline = RawInputsPipeline(
-            DataPipelineParams(mode=PipelineMode.PREDICTION),
+            self.params.pipeline,
             self._data,
             DataGeneratorParams(),
         )
+
         return self.predict_pipeline(pipeline)
 
     def predict_pipeline(self, pipeline: DataPipeline) -> Iterable[Sample]:
diff --git a/tfaip/predict/raw_predictor.py b/tfaip/predict/raw_predictor.py
new file mode 100644
index 0000000..b6553f2
--- /dev/null
+++ b/tfaip/predict/raw_predictor.py
@@ -0,0 +1,68 @@
+import logging
+from contextlib import ExitStack
+from queue import Queue
+from threading import Thread
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from tfaip.predict.predictorbase import PredictorBase
+
+logger = logging.getLogger(__name__)
+
+
+class StopSignal:
+    ...
+
+
+class RawPredictorThread(Thread):
+    def __init__(self, in_queue: Queue, out_queue: Queue, predictor: "PredictorBase"):
+        super().__init__(daemon=True)
+        self.predictor = predictor
+        self.in_queue = in_queue
+        self.out_queue = out_queue
+
+    def run(self) -> None:
+        def generator():
+            while True:
+                sample = self.in_queue.get()
+                if isinstance(sample, StopSignal):
+                    break
+                yield sample
+
+        for sample in self.predictor.predict_raw(generator(), size=5):
+            self.out_queue.put(sample)
+
+
+class RawPredictorCaller:
+    def __init__(self, in_queue, out_queue):
+        self.in_queue = in_queue
+        self.out_queue = out_queue
+
+    def __call__(self, sample):
+        self.in_queue.put(sample)
+        return self.out_queue.get()
+
+
+class RawPredictor:
+    """Utility class to allow for raw prediction but keeping the internal queues open."""
+
+    def __init__(self, predictor: "PredictorBase"):
+        self.predictor = predictor
+        if self.predictor.params.pipeline.batch_size != 1:
+            logger.warning(f"Raw prediction via threading requires batch size == 1. Automatically setting.")
+            self.predictor.params.pipeline.batch_size = 1
+        if not self.predictor.params.silent:
+            logger.warning(f"Consider setting predictor to silent by setting predictor_params.silent = True.")
+
+        self.exit_stack = ExitStack()
+        self.in_queue = Queue(10)
+        self.out_queue = Queue(10)
+        self.thread = RawPredictorThread(self.in_queue, self.out_queue, predictor)
+
+    def __enter__(self):
+        self.thread.start()
+        return RawPredictorCaller(self.in_queue, self.out_queue)
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        self.in_queue.put(StopSignal())
+        self.thread.join()
diff --git a/tfaip/version.py b/tfaip/version.py
index 776a2ac..0a5634c 100644
--- a/tfaip/version.py
+++ b/tfaip/version.py
@@ -15,4 +15,4 @@
 # You should have received a copy of the GNU General Public License along with
 # tfaip. If not, see http://www.gnu.org/licenses/.
 # ==============================================================================
-__version__ = "1.2.1"
+__version__ = "1.2.2"