diff --git a/orchestrators/imajin-pipeline/src/image_pipeline/stages/generate.py b/orchestrators/imajin-pipeline/src/image_pipeline/stages/generate.py
index 656affb3..afec0b34 100644
--- a/orchestrators/imajin-pipeline/src/image_pipeline/stages/generate.py
+++ b/orchestrators/imajin-pipeline/src/image_pipeline/stages/generate.py
@@ -343,6 +343,8 @@ async def init_gpu_boss():
     Must be called at service startup before loading any models.
     Initializes both the GPUBoss for lease coordination and ManagedModelLoader
     for coordinated model loading.
+
+    Automatically detects GPUs and registers them with the coordinator.
     """
     global _boss, _diffusers_loader
 
@@ -351,12 +353,30 @@ async def init_gpu_boss():
         return
 
     from model_boss import GPUBoss, Priority
+    from model_boss.gpu.utils import get_gpu_info_safe
     from model_boss_loaders import ManagedModelLoader, DiffusersLoader  # noqa: F401 - import for side-effect (registry registration)
 
     _boss = GPUBoss()
     await _boss.connect()
     logger.info("GPU coordination initialized (model-boss v3)")
 
+    # Detect and initialize GPUs
+    gpus = await get_gpu_info_safe(timeout_s=10.0)
+    if not gpus:
+        raise RuntimeError(
+            "No GPUs detected. GPU coordination requires at least one CUDA-capable GPU."
+        )
+
+    for gpu in gpus:
+        await _boss.initialize_gpu(
+            gpu_index=gpu["index"],
+            vram_total_mb=gpu["total_memory_mb"],
+            gpu_name=gpu["name"],
+        )
+        logger.info(
+            f"Registered GPU {gpu['index']}: {gpu['name']} ({gpu['total_memory_mb']} MB)"
+        )
+
     # Initialize managed loader with GPU boss for coordinated loading
     _diffusers_loader = ManagedModelLoader(
         boss=_boss,