diff --git a/pyproject.toml b/pyproject.toml
index 6947df8..d0d776a 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -42,7 +42,9 @@ dependencies = [
     "matplotlib",
     "mrcfile",
     "plotille",
-    "rich"
+    "rich",
+    "starfile",
+    "pydantic>=2.0"
 ]
 
 # https://peps.python.org/pep-0621/#dependencies-optional-dependencies
diff --git a/src/ttfsc/_cli.py b/src/ttfsc/_cli.py
index 2845987..ac5da1e 100644
--- a/src/ttfsc/_cli.py
+++ b/src/ttfsc/_cli.py
@@ -97,6 +97,7 @@ def ttfsc_cli(
                 estimated_resolution_frequency_pixel,
                 correction_from_resolution_angstrom,
                 fsc_values_corrected,
+                fsc_values_randomized,
             ) = calculate_noise_injected_fsc(
                 map1_tensor,
                 map2_tensor,
@@ -113,6 +114,45 @@ def ttfsc_cli(
                 f"criterion with correction after {correction_from_resolution_angstrom:.2f} Å: "
                 f"{estimated_resolution_angstrom:.2f} Å"
             )
+    if save_starfile:
+        import pandas as pd
+        import starfile
+        from numpy import nan
+
+        from ._starfile_schema import RelionDataGeneral, RelionFSCData
+
+        data_general = RelionDataGeneral(
+            rlnFinalResolution=estimated_resolution_angstrom,
+            rlnUnfilteredMapHalf1=map1.name,
+            rlnUnfilteredMapHalf2=map2.name,
+        )
+        if mask != Masking.none:
+            data_general.rlnParticleBoxFractionSolventMask = mask_tensor.sum().item() / mask_tensor.numel()
+        if correct_for_masking:
+            data_general.rlnRandomiseFrom = correction_from_resolution_angstrom
+
+        fsc_data = []
+        for i, (f, r) in enumerate(zip(fsc_values_unmasked, resolution_angstroms)):
+            fsc_data.append(
+                RelionFSCData(
+                    rlnSpectralIndex=i,
+                    rlnResolution=r,
+                    rlnAngstromResolution=r,
+                    rlnFourierShellCorrelationCorrected=fsc_values_corrected[i] if correct_for_masking else nan,
+                    rlnFourierShellCorrelationUnmaskedMaps=f,
+                    rlnFourierShellCorrelationMaskedMaps=fsc_values_masked[i] if fsc_values_masked is not None else nan,
+                    rlnCorrectedFourierShellCorrelationPhaseRandomizedMaskedMaps=fsc_values_randomized[i]
+                    if correct_for_masking
+                    else nan,
+                    rlnFourierShellCorrelationParticleMaskFraction=1
+                    - ((1 - f) * data_general.rlnParticleBoxFractionSolventMask)
+                    if mask != Masking.none
+                    else nan,
+                )
+            )
+        starfile.write(
+            {"general": data_general.__dict__, "fsc": pd.DataFrame([f.__dict__ for f in fsc_data])}, save_starfile
+        )
 
     if plot:
         from ._plotting import plot_matplotlib, plot_plottile
diff --git a/src/ttfsc/_masking.py b/src/ttfsc/_masking.py
index afe9a2d..ce41950 100644
--- a/src/ttfsc/_masking.py
+++ b/src/ttfsc/_masking.py
@@ -20,7 +20,7 @@ def calculate_noise_injected_fsc(
     estimated_resolution_frequency_pixel: float,
     correct_from_resolution: Optional[float] = None,
     correct_from_fraction_of_estimated_resolution: float = 0.5,
-):
+) -> tuple[float, float, float, torch.tensor, torch.tensor]:
     from torch_grid_utils import fftfreq_grid
 
     map1_tensor_randomized = torch.fft.rfftn(map1_tensor)
@@ -75,6 +75,7 @@ def calculate_noise_injected_fsc(
         estimated_resolution_frequency_pixel,
         correct_from_resolution,
         fsc_values_corrected,
+        fsc_values_masked_randomized,
     )
 
 
@@ -106,7 +107,7 @@ def calculate_masked_fsc(
         mask_tensor[inside_sphere] = 1
 
         # if requested, a soft edge is added to the mask
-        mask_tensor = add_soft_edge(mask_tensor, mask_soft_edge_width_pixels)
+        mask_tensor = torch.tensor(add_soft_edge(mask_tensor, mask_soft_edge_width_pixels))
 
         map1_tensor_masked = map1_tensor * mask_tensor
         map2_tensor_masked = map2_tensor * mask_tensor
diff --git a/src/ttfsc/_starfile_schema.py b/src/ttfsc/_starfile_schema.py
new file mode 100644
index 0000000..e3b9215
--- /dev/null
+++ b/src/ttfsc/_starfile_schema.py
@@ -0,0 +1,29 @@
+from typing import List
+
+from numpy import nan
+from pydantic import BaseModel
+
+
+class RelionDataGeneral(BaseModel):
+    rlnFinalResolution: float
+    rlnUnfilteredMapHalf1: str
+    rlnUnfilteredMapHalf2: str
+    rlnParticleBoxFractionSolventMask: float = nan
+    rlnRandomiseFrom: float = nan
+    rlnMaskName: str = ""
+
+
+class RelionFSCData(BaseModel):
+    rlnSpectralIndex: int
+    rlnResolution: float
+    rlnAngstromResolution: float
+    rlnFourierShellCorrelationCorrected: float
+    rlnFourierShellCorrelationParticleMaskFraction: float
+    rlnFourierShellCorrelationUnmaskedMaps: float
+    rlnFourierShellCorrelationMaskedMaps: float
+    rlnCorrectedFourierShellCorrelationPhaseRandomizedMaskedMaps: float
+
+
+class RelionStarfile(BaseModel):
+    data_general: RelionDataGeneral
+    fsc_data: List[RelionFSCData]