From 5727c391892f2bbae69512eba32488d84e1edd1c Mon Sep 17 00:00:00 2001
From: digital-garbage <133010426+digital-garbage@users.noreply.github.com>
Date: Sun, 19 Jul 2026 00:53:36 +0300
Subject: [PATCH] Fix CLIPSeg multi-prompt crash: missing padding/truncation on
 tokenizer call

Live rental run hit it immediately on 'detail_targets' with more than one
target: CLIPSegProcessor tokenizes each prompt to a different length, and
without padding=True torch can't stack them into a batch ('excessive
nesting... type list where int is expected'). Reproduced against the real
CIDAS/clipseg-rd64-refined model before and after the fix. Single-target
runs were unaffected (batch of 1 never needed padding), which is why the
earlier probe steps didn't catch it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 detailing.py            |  7 ++++++-
 tests/test_detailing.py | 35 +++++++++++++++++++++++++++++++++++
 2 files changed, 41 insertions(+), 1 deletion(-)

diff --git a/detailing.py b/detailing.py
index 96adea14..43918672 100644
--- a/detailing.py
+++ b/detailing.py
@@ -224,8 +224,13 @@ def _clipseg_heat(frames, prompts, debug=False):
     with torch.no_grad():
         for _, img in frames:
             pil = Image.fromarray((img.numpy() * 255.0).astype(np.uint8))
+            # padding=True is mandatory once len(prompts) > 1: CLIPSeg's tokenizer
+            # otherwise returns one variable-length id list per prompt and torch can't
+            # stack them into a batch ("excessive nesting... type list where int is
+            # expected"). truncation=True guards the rare very-long user-typed target.
             inputs = processor(
-                text=list(prompts), images=[pil] * len(prompts), return_tensors="pt")
+                text=list(prompts), images=[pil] * len(prompts), return_tensors="pt",
+                padding=True, truncation=True)
             logits = model(**inputs).logits  # [P, 352, 352] (or [352, 352] for one prompt)
             if logits.dim() == 2:
                 logits = logits[None]
diff --git a/tests/test_detailing.py b/tests/test_detailing.py
index 1430600f..b3705287 100644
--- a/tests/test_detailing.py
+++ b/tests/test_detailing.py
@@ -99,6 +99,41 @@ def test_find_tube_enforces_minimum_edge():
     assert (x1 - x0) >= detailing.MIN_TUBE_EDGE
 
 
+# ---------------------------------------------------------------------------
+# _clipseg_heat
+# ---------------------------------------------------------------------------
+
+def test_clipseg_heat_passes_padding_and_truncation(monkeypatch):
+    """Regression: a multi-prompt call without padding=True crashes CLIPSeg's
+    tokenizer ("excessive nesting... type list where int is expected") the moment
+    detail_targets has more than one entry — caught live on a real rental run.
+    truncation=True is the same call, cheap to require alongside it."""
+    calls = []
+
+    class _FakeInputs(dict):
+        def to(self, *a, **k):
+            return self
+
+    class _FakeProcessor:
+        def __call__(self, text, images, return_tensors, **kwargs):
+            calls.append(kwargs)
+            return _FakeInputs()
+
+    class _FakeModel:
+        def __call__(self, **inputs):
+            n = 2
+            return types.SimpleNamespace(logits=torch.zeros(n, 8, 8))
+
+        def eval(self):
+            return self
+
+    monkeypatch.setattr(detailing, "_get_clipseg", lambda: (_FakeProcessor(), _FakeModel()))
+    frames = [(0, torch.zeros(16, 16, 3))]
+    detailing._clipseg_heat(frames, ["hands", "feet"])
+    assert calls and calls[0].get("padding") is True
+    assert calls[0].get("truncation") is True
+
+
 # ---------------------------------------------------------------------------
 # resolve_upsampler_name
 # ---------------------------------------------------------------------------