diff --git a/modules/image_models.py b/modules/image_models.py
index 2ef1e730..8363533b 100644
--- a/modules/image_models.py
+++ b/modules/image_models.py
@@ -98,7 +98,7 @@ def load_image_model(model_name, dtype='bfloat16', attn_backend='sdpa', cpu_offl
     Args:
         model_name: Name of the model directory
         dtype: 'bfloat16' or 'float16'
-        attn_backend: 'sdpa', 'flash_attention_2', or 'flash_attention_3'
+        attn_backend: 'sdpa' or 'flash_attention_2'
         cpu_offload: Enable CPU offloading for low VRAM
         compile_model: Compile the model for faster inference (slow first run)
         quant_method: 'none', 'bnb-8bit', 'bnb-4bit', or torchao options (int8wo, fp4, float8wo)
@@ -145,8 +145,6 @@ def load_image_model(model_name, dtype='bfloat16', attn_backend='sdpa', cpu_offl
         if hasattr(pipe, 'transformer') and hasattr(pipe.transformer, 'set_attention_backend'):
             if attn_backend == 'flash_attention_2':
                 pipe.transformer.set_attention_backend("flash")
-            elif attn_backend == 'flash_attention_3':
-                pipe.transformer.set_attention_backend("_flash_3")
             # sdpa is the default, no action needed
 
         if compile_model:
diff --git a/modules/shared.py b/modules/shared.py
index 4a0fd986..4e17497b 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -56,7 +56,7 @@ group = parser.add_argument_group('Image model')
 group.add_argument('--image-model', type=str, help='Name of the image model to select on startup (overrides saved setting).')
 group.add_argument('--image-model-dir', type=str, default='user_data/image_models', help='Path to directory with all the image models.')
 group.add_argument('--image-dtype', type=str, default=None, choices=['bfloat16', 'float16'], help='Data type for image model.')
-group.add_argument('--image-attn-backend', type=str, default=None, choices=['sdpa', 'flash_attention_2', 'flash_attention_3'], help='Attention backend for image model.')
+group.add_argument('--image-attn-backend', type=str, default=None, choices=['sdpa', 'flash_attention_2'], help='Attention backend for image model.')
 group.add_argument('--image-cpu-offload', action='store_true', help='Enable CPU offloading for image model.')
 group.add_argument('--image-compile', action='store_true', help='Compile the image model for faster inference.')
 group.add_argument('--image-quant', type=str, default=None,
diff --git a/modules/ui_image_generation.py b/modules/ui_image_generation.py
index 08cf3f64..fdf1af86 100644
--- a/modules/ui_image_generation.py
+++ b/modules/ui_image_generation.py
@@ -485,7 +485,7 @@ def create_ui():
                                     info='bfloat16 recommended for modern GPUs'
                                 )
                                 shared.gradio['image_attn_backend'] = gr.Dropdown(
-                                    choices=['sdpa', 'flash_attention_2', 'flash_attention_3'],
+                                    choices=['sdpa', 'flash_attention_2'],
                                     value=shared.settings['image_attn_backend'],
                                     label='Attention Backend',
                                     info='SDPA is default. Flash Attention requires compatible GPU.'