Use named character refs for H3 image conditioning

dumas_h3_longvideos.py

@@ -4501,11 +4501,31 @@ def resolve_tagged_refs(text, ref_list):
         return f"<Picture {renumber[n]}>" if n in renumber else ""
 
     out = _PICTURE_TAG.sub(sub, text or "")
-    if dropped:                       # tidy the gap a removed tag leaves behind
+    if dropped:                       # tidy the gap a removed tag leaves behind
-        out = re.sub(r"\s+([,.;:])", r"\1", out)
+        out = re.sub(r"\s+([,.;:])", r"\1", out)
-        out = re.sub(r"(,\s*){2,}", ", ", out)
+        out = re.sub(r"(,\s*){2,}", ", ", out)
-        out = re.sub(r"\s{2,}", " ", out)
+        out = re.sub(r"\s{2,}", " ", out)
-    return out.strip(), [ref_list[n - 1] for n in live], dropped
+    return out.strip(), [ref_list[n - 1] for n in live], dropped
+
+
+def resolve_prompt_refs(text, ref_list):
+    """(rewritten text, refs, dropped) for the refs a shot actually carries.
+
+    Explicit <Picture N> tags still decide which slot numbers the prompt points at,
+    but named character matches must ride into the real ref-image list too. Without
+    that split, a shot could inherit the facts/context for "Mara" and "Jon" while
+    only carrying a tagged location image, which reads exactly like the names were
+    understood but the faces were ignored."""
+    rewritten, tagged_refs, dropped = resolve_tagged_refs(text, ref_list)
+    refs = list(tagged_refs)
+    seen = {id(ref) for ref in refs}
+    for _slot_number, ref in _named_character_refs_for_text(rewritten, ref_list):
+        marker = id(ref)
+        if marker in seen:
+            continue
+        seen.add(marker)
+        refs.append(ref)
+    return rewritten, refs, dropped
 
 
 def shot_references(ref_list, ref_mode, shot_index, handoff):
@@ -6621,7 +6641,7 @@ class H3LongVideos:
                 for shot_index, gen in enumerate(gens):
                     shot_mode = beat_ref_mode_directive(beats[shot_index] if shot_index < len(beats) else "") or ref_mode
                     if shot_mode in ("where tagged", "auto ref2v") and any_tags_anywhere:
-                        if resolve_tagged_refs(gen, ref_slots)[1]:
+                        if resolve_prompt_refs(gen, ref_slots)[1]:
                             on.append(shot_index + 1)
                             tagged_used = True
                     else:
@@ -6715,7 +6735,7 @@ class H3LongVideos:
                 # The prompt itself says where each reference belongs: the shot whose
                 # text names <Picture N> gets image N, renumbered to match what that
                 # shot actually carries. Every untagged shot keeps its handoff.
-                gen_prompt, shot_refs, dropped = resolve_tagged_refs(gen_prompt, ref_slots)
+                gen_prompt, shot_refs, dropped = resolve_prompt_refs(gen_prompt, ref_slots)
                 for n in dropped:
                     if n not in ref_missing:
                         ref_missing.append(n)

tests/test_dumas_h3_longvideos.py

@@ -388,6 +388,25 @@ class DumasH3LongVideosHelperTests(unittest.TestCase):
         )
         self.assertEqual(dropped, [4])
 
+    def test_resolve_prompt_refs_keeps_named_character_images_alongside_tagged_location(self):
+        refs = [
+            {"kind": "character", "image": "img1", "name": "Mara"},
+            {"kind": "character", "image": "img2", "name": "Jon"},
+            {"kind": "location", "image": "img3", "name": "Hangar"},
+        ]
+
+        text, references, dropped = self.module.resolve_prompt_refs(
+            "Mara and Jon argue inside <Picture 3>.",
+            refs,
+        )
+
+        self.assertEqual(text, "Mara and Jon argue inside <Picture 1>.")
+        self.assertEqual(
+            [self.module._reference_image(ref) for ref in references],
+            ["img3", "img1", "img2"],
+        )
+        self.assertEqual(dropped, [])
+
     def test_shot_references_uses_all_connected_sparse_slots(self):
         refs = [
             None,