diff --git a/src/maxdiffusion/models/embeddings_flax.py b/src/maxdiffusion/models/embeddings_flax.py
index 17ac2b07a..e51308213 100644
--- a/src/maxdiffusion/models/embeddings_flax.py
+++ b/src/maxdiffusion/models/embeddings_flax.py
@@ -546,17 +546,21 @@ class CombinedTimestepGuidanceTextProjEmbeddings(nn.Module):
 
   @nn.compact
   def __call__(self, timestep, guidance, pooled_projection=None):
-    timesteps_proj = FlaxTimesteps(dim=self.frequency_embedding_size, flip_sin_to_cos=True, freq_shift=0)(timestep)
-    dtype = pooled_projection.dtype if pooled_projection is not None else jnp.float32
+    # timestep and guidance arrive already projected: FluxTransformer2DModel runs
+    # both through its own timestep_embedding(), which carries a time_factor the
+    # sinusoidal helper here does not. Projecting again would shape them
+    # (batch, frequency_embedding_size, embedding_dim) and break the sum with the
+    # pooled projection below. The NNX variant owns its projection instead.
+    timesteps_proj = timestep
     timestep_emb = FlaxTimestepEmbedding(
         time_embed_dim=self.embedding_dim, dtype=self.dtype, weights_dtype=self.weights_dtype
-    )(timesteps_proj.astype(dtype))
+    )(timesteps_proj)
 
     if self.guidance_embeds and guidance is not None:
-      guidance_proj = FlaxTimesteps(dim=self.frequency_embedding_size, flip_sin_to_cos=True, freq_shift=0)(guidance)
+      guidance_proj = guidance
       guidance_emb = FlaxTimestepEmbedding(
           time_embed_dim=self.embedding_dim, dtype=self.dtype, weights_dtype=self.weights_dtype
-      )(guidance_proj.astype(dtype))
+      )(guidance_proj)
       time_guidance_emb = timestep_emb + guidance_emb
     else:
       time_guidance_emb = timestep_emb
diff --git a/src/maxdiffusion/models/flux/transformers/transformer_flux_flax.py b/src/maxdiffusion/models/flux/transformers/transformer_flux_flax.py
index 2d536ca1c..026aaded6 100644
--- a/src/maxdiffusion/models/flux/transformers/transformer_flux_flax.py
+++ b/src/maxdiffusion/models/flux/transformers/transformer_flux_flax.py
@@ -545,6 +545,9 @@ def setup(self):
         self.inner_dim,
         elementwise_affine=False,
         eps=self.eps,
+        # FLUX loads adaLN_modulation_1 from the original checkpoint, which emits
+        # shift before scale. The module default follows Flux.2-Klein instead.
+        scale_shift_order="shift_scale",
         dtype=self.dtype,
         weights_dtype=self.weights_dtype,
         precision=self.precision,
diff --git a/src/maxdiffusion/models/flux/util.py b/src/maxdiffusion/models/flux/util.py
index 77e34099c..f651368ec 100644
--- a/src/maxdiffusion/models/flux/util.py
+++ b/src/maxdiffusion/models/flux/util.py
@@ -260,7 +260,7 @@ def load_flow_model(name: str, eval_shapes: dict, device: str, hf_download: bool
           renamed_pt_key = renamed_pt_key.replace("out_layer", "linear_2")
         elif "final_layer" in renamed_pt_key:
           renamed_pt_key = renamed_pt_key.replace("final_layer.linear", "proj_out")
-          renamed_pt_key = renamed_pt_key.replace("final_layer.adaLN_modulation_1", "norm_out.Dense_0")
+          renamed_pt_key = renamed_pt_key.replace("final_layer.adaLN_modulation_1", "norm_out.linear")
 
         pt_tuple_key = tuple(renamed_pt_key.split("."))
         flax_key, flax_tensor = rename_key_and_reshape_tensor(pt_tuple_key, tensor, eval_shapes)
diff --git a/src/maxdiffusion/tests/flux_conditioning_test.py b/src/maxdiffusion/tests/flux_conditioning_test.py
new file mode 100644
index 000000000..813f17840
--- /dev/null
+++ b/src/maxdiffusion/tests/flux_conditioning_test.py
@@ -0,0 +1,95 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for the FLUX conditioning path (CPU backend).
+
+FLUX shares CombinedTimestepGuidanceTextProjEmbeddings and
+AdaLayerNormContinuous with Flux.2-Klein, which reads them with different
+conventions. These cover the conventions FLUX needs.
+"""
+
+import os
+import unittest
+
+os.environ.setdefault("JAX_PLATFORMS", "cpu")
+
+import flax
+import jax
+import jax.numpy as jnp
+import numpy as np
+
+from maxdiffusion.models.embeddings_flax import CombinedTimestepGuidanceTextProjEmbeddings
+from maxdiffusion.models.normalization_flax import AdaLayerNormContinuous
+
+
+def _layer_norm(x, eps=1e-5):
+  mean = np.mean(x, axis=-1, keepdims=True)
+  variance = np.var(x, axis=-1, keepdims=True)
+  return (x - mean) / np.sqrt(variance + eps)
+
+
+class FluxConditioningTest(unittest.TestCase):
+
+  def test_guidance_embeddings_accept_projected_timesteps(self):
+    """FluxTransformer2DModel projects timestep and guidance before this module."""
+    batch, frequency_embedding_size, embedding_dim, pooled_projection_dim = 2, 256, 64, 32
+    module = CombinedTimestepGuidanceTextProjEmbeddings(
+        embedding_dim=embedding_dim, pooled_projection_dim=pooled_projection_dim
+    )
+    timestep = jnp.zeros((batch, frequency_embedding_size), jnp.float32)
+    guidance = jnp.zeros((batch, frequency_embedding_size), jnp.float32)
+    pooled_projection = jnp.zeros((batch, pooled_projection_dim), jnp.float32)
+
+    variables = module.init(jax.random.PRNGKey(0), timestep, guidance, pooled_projection)
+    conditioning = module.apply(variables, timestep, guidance, pooled_projection)
+
+    self.assertEqual(conditioning.shape, (batch, embedding_dim))
+
+  def test_norm_out_dense_is_named_linear(self):
+    """load_flow_model writes final_layer.adaLN_modulation_1 to norm_out/linear."""
+    module = AdaLayerNormContinuous(embedding_dim=4, elementwise_affine=False)
+    x = jnp.zeros((2, 3, 4), jnp.float32)
+    conditioning_embedding = jnp.zeros((2, 8), jnp.float32)
+
+    variables = module.init(jax.random.PRNGKey(0), x, conditioning_embedding)
+
+    self.assertIn("linear", variables["params"])
+
+  def test_shift_scale_order_reads_shift_first(self):
+    """The original FLUX checkpoint emits shift ahead of scale."""
+    embedding_dim, conditioning_dim, batch, sequence = 4, 8, 2, 3
+    x = jax.random.normal(jax.random.PRNGKey(1), (batch, sequence, embedding_dim))
+    conditioning_embedding = jax.random.normal(jax.random.PRNGKey(2), (batch, conditioning_dim))
+
+    def modulate(scale_shift_order):
+      module = AdaLayerNormContinuous(
+          embedding_dim=embedding_dim, elementwise_affine=False, scale_shift_order=scale_shift_order
+      )
+      variables = flax.linen.meta.unbox(flax.core.unfreeze(module.init(jax.random.PRNGKey(0), x, conditioning_embedding)))
+      # A zero kernel leaves the bias as the whole modulation, so the halves are
+      # known: 2.0 in the first, 0.0 in the second.
+      variables["params"]["linear"]["kernel"] = jnp.zeros_like(variables["params"]["linear"]["kernel"])
+      variables["params"]["linear"]["bias"] = jnp.concatenate([jnp.full((embedding_dim,), 2.0), jnp.zeros((embedding_dim,))])
+      return module.apply(variables, x, conditioning_embedding)
+
+    normalized = _layer_norm(np.asarray(x))
+
+    # shift first: (1 + 0) * norm + 2
+    np.testing.assert_allclose(np.asarray(modulate("shift_scale")), normalized + 2.0, atol=1e-4)
+    # scale first: (1 + 2) * norm + 0
+    np.testing.assert_allclose(np.asarray(modulate("scale_shift")), 3.0 * normalized, atol=1e-4)
+
+
+if __name__ == "__main__":
+  unittest.main()