diff --git a/services/imajin-diffusion/service/src/api/main.py b/services/imajin-diffusion/service/src/api/main.py
index d115d066..b2a04957 100644
--- a/services/imajin-diffusion/service/src/api/main.py
+++ b/services/imajin-diffusion/service/src/api/main.py
@@ -81,25 +81,18 @@ async def init_service() -> None:
     """
     logger.info("Image Generation Service starting...")
 
-    # Shared GPU lock — serialises GenerationQueue and BackgroundInpainter so
-    # they never run simultaneously and compete for VRAM.
-    gpu_lock = asyncio.Lock()
-    lifespan.set_state("gpu_lock", gpu_lock)
-
-    # Generation queue — manages GPU lease lifecycle per-session
+    # Generation queue — serialises pipeline execution; model-boss handles GPU leasing
     import os
-    vram_mb = int(os.environ.get("PIPELINE_VRAM_MB", "16000"))
     idle_timeout = float(os.environ.get("PIPELINE_IDLE_TIMEOUT_SECONDS", "300"))
-    queue = GenerationQueue(vram_mb=vram_mb, idle_timeout_seconds=idle_timeout, gpu_lock=gpu_lock)
+    queue = GenerationQueue(idle_timeout_seconds=idle_timeout)
     await queue.start()
     lifespan.set_state("generation_queue", queue)
-    logger.info(
-        "GenerationQueue started (vram=%dMB, idle_timeout=%.0fs)",
-        vram_mb, idle_timeout,
-    )
+    logger.info("GenerationQueue started (idle_timeout=%.0fs)", idle_timeout)
 
     # Background inpainter — SDXL inpainting for /repaint-background/async
-    background_inpainter = BackgroundInpainter(gpu_lock=gpu_lock)
+    # Uses its own lock to serialise concurrent inpainting requests within this service.
+    inpainter_lock = asyncio.Lock()
+    background_inpainter = BackgroundInpainter(gpu_lock=inpainter_lock)
     lifespan.set_state("background_inpainter", background_inpainter)
     logger.info("BackgroundInpainter registered (model loaded on first use)")