Revert "Revert long videos ref routing to release"

This reverts commit 00985b73b3.
This commit is contained in:
2026-08-30 21:37:15 +00:00
parent b7c3bba199
commit a2dabaab7e
2 changed files with 93 additions and 8 deletions
+28 -8
View File
@@ -4558,7 +4558,7 @@ def _connected_refs(ref_slots):
return [ref for ref in (ref_slots or []) if _reference_image(ref) is not None]
def resolve_tagged_refs(text, ref_list):
def resolve_tagged_refs(text, ref_list):
"""(rewritten text, images, dropped) for the <Picture N> tags in ONE shot.
The tokenizer numbers references by their position in the list it is handed, so
@@ -4579,11 +4579,31 @@ 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
out = re.sub(r"\s+([,.;:])", r"\1", 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
if dropped: # tidy the gap a removed tag leaves behind
out = re.sub(r"\s+([,.;:])", r"\1", 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
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):
@@ -6750,7 +6770,7 @@ 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:
@@ -6845,7 +6865,7 @@ 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)
+65
View File
@@ -423,6 +423,41 @@ 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_resolve_prompt_refs_where_tagged_mode_stays_tag_only(self):
refs = [
{"kind": "character", "image": "img1", "name": "Mara"},
{"kind": "location", "image": "img2", "name": "Hangar"},
]
text, references, dropped = self.module.resolve_prompt_refs(
"Mara waits in the hangar near <Picture 2>.",
refs,
include_named=False,
)
self.assertEqual(text, "Mara waits in the hangar near <Picture 1>.")
self.assertEqual([self.module._reference_image(ref) for ref in references], ["img2"])
self.assertEqual(dropped, [])
def test_resolve_shot_references_uses_named_characters_without_picture_tags(self):
refs = [
{"kind": "character", "image": "img1", "name": "Mara"},
@@ -444,6 +479,21 @@ class DumasH3LongVideosHelperTests(unittest.TestCase):
self.assertFalse(shot_tag_driven)
self.assertEqual(mode_eff, "auto ref2v")
def test_resolve_prompt_refs_prioritizes_characters_before_locations(self):
refs = [
{"kind": "location", "image": "img1", "name": "Hangar"},
{"kind": "character", "image": "img2", "name": "Mara"},
]
text, references, dropped = self.module.resolve_prompt_refs(
"[Generation 1] Mara waits in the hangar.",
refs,
)
self.assertEqual(text, "[Generation 1] Mara waits in the hangar.")
self.assertEqual([self.module._reference_image(ref) for ref in references], ["img2", "img1"])
self.assertEqual(dropped, [])
def test_shot_references_uses_all_connected_sparse_slots(self):
refs = [
None,
@@ -697,6 +747,21 @@ class DumasH3LongVideosHelperTests(unittest.TestCase):
self.assertIn("Persistent appearance for Mara: silver hair.", context)
self.assertNotIn("Persistent wardrobe/style for Mara: red jacket.", context)
def test_resolve_prompt_refs_adds_named_location_refs(self):
refs = [
{"kind": "location", "image": "img2", "name": "Hangar", "description": "wet concrete floor"},
]
rewritten, matched, dropped = self.module.resolve_prompt_refs(
"[Generation 1] They wait in the hangar.",
refs,
)
self.assertEqual(rewritten, "[Generation 1] They wait in the hangar.")
self.assertEqual(dropped, [])
self.assertEqual(len(matched), 1)
self.assertEqual(matched[0]["name"], "Hangar")
def test_resolve_tagged_refs_drops_reference_without_image(self):
refs = [
{"kind": "character", "image": None, "name": "Mara"},