fix(ai): drop per-frame pymatting in animated background removal (#684)

The frame loop spent 2-19s per frame in CPU pymatting while the CUDA session sat idle; a 30-frame GIF took 353s on a 4070. Animations skip alpha matting now (stills keep it), the session device is logged, and the CUDA-to-CPU session fallback says why. Same fixture finishes in 46s with every frame on the GPU. Fixes #668.

packages/ai/python/gif_remove_bg.py

@@ -204,8 +204,13 @@ def _create_session(model, providers, device):
     _register_hr_matting_session(sessions_class)
     try:
         return new_session(model, providers=providers), device
-    except Exception:
+    except Exception as e:
         if "CUDAExecutionProvider" in providers:
+            print(
+                f"[gif-remove-bg] CUDA session for '{model}' failed ({e}); falling back to CPU",
+                file=sys.stderr,
+                flush=True,
+            )
             return new_session(model, providers=["CPUExecutionProvider"]), "cpu"
         raise
 
@@ -291,11 +296,21 @@ def main():
         emit_progress(5, "Loading model")
         providers, device = onnx_providers()
         session, device = _create_session(model, providers, device)
-        use_alpha = device != "cpu"
+        # Per-frame pymatting runs on the CPU whatever the session device and
+        # costs seconds per frame (it dominated a 30-frame GIF at ~12s/frame
+        # on an otherwise idle GPU host, #668), and its per-frame noise
+        # flickers across an animation. Stills keep matting; animations never
+        # use it.
+        use_alpha = False
+        print(
+            f"[gif-remove-bg] session device={device} model={model}",
+            file=sys.stderr,
+            flush=True,
+        )
 
-        # Probe frame 0 to settle model (OOM -> lighter model for the WHOLE
-        # animation, never per-frame) and matting viability once. Switching model
-        # or matting mid-animation would flicker.
+        # Probe frame 0 to settle the model once (OOM -> lighter model for the
+        # WHOLE animation, never per-frame; switching mid-animation would
+        # flicker).
         try:
             _remove_one(frames[0], session, use_alpha, settings, target, orig_size)
         except Exception as e:
@@ -303,9 +318,6 @@ def main():
                 model = "u2net"
                 emit_progress(5, "Retrying with a lighter model")
                 session, device = _create_session(model, providers, device)
-                use_alpha = device != "cpu"
-            elif use_alpha:
-                use_alpha = False  # matting not viable on this device/model
             else:
                 raise
 

packages/ai/python/tests/test_gif_remove_bg_session.py

@@ -0,0 +1,77 @@
+"""The animated-removal session helper must not downgrade to CPU silently.
+
+The 2.2.0 fleet QA (#668) burned hours on an unexplainable six-minute GIF job
+because nothing in the logs said which device the frame loop used, and the
+CUDA-to-CPU retry in _create_session swallowed the reason entirely.
+"""
+
+import importlib.util
+import os
+import sys
+import types
+
+
+def load_gif_module(fake_new_session):
+    """Load gif_remove_bg with fakes for its function-local imports."""
+    rembg = types.ModuleType("rembg")
+    rembg.new_session = fake_new_session
+    rembg.remove = lambda *a, **k: b""
+    sessions = types.ModuleType("rembg.sessions")
+    sessions.sessions_class = []
+    rembg.sessions = sessions
+
+    remove_bg = types.ModuleType("remove_bg")
+    remove_bg._register_matting_session = lambda _c: None
+    remove_bg._register_hr_matting_session = lambda _c: None
+    remove_bg.ALLOWED_MODELS = {"u2net"}
+
+    sys.modules["rembg"] = rembg
+    sys.modules["rembg.sessions"] = sessions
+    sys.modules["remove_bg"] = remove_bg
+
+    script_path = os.path.join(os.path.dirname(__file__), "..", "gif_remove_bg.py")
+    spec = importlib.util.spec_from_file_location("gif_remove_bg_under_test", script_path)
+    module = importlib.util.module_from_spec(spec)
+    assert spec.loader is not None
+    spec.loader.exec_module(module)
+    return module
+
+
+def teardown_function(_fn):
+    for name in ("rembg", "rembg.sessions", "remove_bg"):
+        sys.modules.pop(name, None)
+
+
+def test_cuda_session_failure_falls_back_to_cpu_and_says_so(capsys):
+    calls = []
+
+    def new_session(model, providers=None):
+        calls.append(providers)
+        if "CUDAExecutionProvider" in providers:
+            raise RuntimeError("CUDA failure for the test")
+        return "cpu-session"
+
+    mod = load_gif_module(new_session)
+    session, device = mod._create_session(
+        "u2net", ["CUDAExecutionProvider", "CPUExecutionProvider"], "cuda"
+    )
+
+    assert session == "cpu-session"
+    assert device == "cpu"
+    assert calls == [
+        ["CUDAExecutionProvider", "CPUExecutionProvider"],
+        ["CPUExecutionProvider"],
+    ]
+    err = capsys.readouterr().err
+    assert "falling back to CPU" in err
+    assert "CUDA failure for the test" in err
+
+
+def test_successful_cuda_session_keeps_device_and_stays_quiet(capsys):
+    mod = load_gif_module(lambda model, providers=None: "gpu-session")
+    session, device = mod._create_session(
+        "u2net", ["CUDAExecutionProvider", "CPUExecutionProvider"], "cuda"
+    )
+    assert session == "gpu-session"
+    assert device == "cuda"
+    assert "falling back" not in capsys.readouterr().err