234 lines
14 KiB
Python
234 lines
14 KiB
Python
"""REF2VA Forge easy mode: labelled pictures, code-written cast, music only when asked."""
|
|
import importlib.util
|
|
import sys
|
|
from pathlib import Path
|
|
|
|
ROOT = Path(__file__).resolve().parents[2]
|
|
sys.path.insert(0, str(ROOT / "nodes"))
|
|
spec = importlib.util.spec_from_file_location("h3_forge_test", ROOT / "nodes" / "h3_forge.py")
|
|
forge = importlib.util.module_from_spec(spec)
|
|
spec.loader.exec_module(forge)
|
|
|
|
|
|
def pic(easy_role, **extra):
|
|
return {"kind": "image", "role": "subject", "easy_role": easy_role, **extra}
|
|
|
|
|
|
REFS = [pic("character-2"), pic("place"), pic("character-1", keep="tired"), pic("character-2"), pic("first-frame")]
|
|
|
|
|
|
def test_cast_numbers_characters_then_place_and_lists_frames_apart():
|
|
cast = forge.easy_cast(REFS)
|
|
assert [(s["tag"], s["name"], s["pictures"]) for s in cast["subjects"]] == [
|
|
("<Subject 1>", "Character 1", [3]), ("<Subject 2>", "Character 2", [1, 4]), ("<Subject 3>", "the place", [2])]
|
|
assert [(f["picture"], f["which"]) for f in cast["frames"]] == [(5, "first")]
|
|
|
|
|
|
def test_group_picture_places_each_character():
|
|
refs = [pic("character-1"), pic("group-21"), pic("place"), pic("group-123")]
|
|
cast = forge.easy_cast(refs)
|
|
by = {s["name"]: s for s in cast["subjects"]}
|
|
assert by["Character 1"]["pictures"] == [1]
|
|
assert by["Character 1"]["placements"] == [{"picture": 2, "position": "right"}, {"picture": 4, "position": "left"}]
|
|
assert by["Character 3"]["tag"] == "<Subject 3>" and by["Character 3"]["pictures"] == []
|
|
assert by["the place"]["tag"] == "<Subject 4>"
|
|
lines = "\n".join(forge.easy_lines(cast))
|
|
assert '<Subject 1> is "Character 1" in the brief, a character, shown in <Picture 1>, and on the right in <Picture 2> and on the left in <Picture 4>' in lines
|
|
assert '<Subject 3> is "Character 3" in the brief, a character, shown on the right in <Picture 4>' in lines
|
|
segments = {"Detailed description": "[Shot 1] <Subject 1>, <Subject 2> and <Subject 3> in <Subject 4>."}
|
|
forge.easy_segments(cast, segments)
|
|
assert "<Subject 2> is the character on the left in <Picture 2> and in the middle in <Picture 4>;" in segments["Subject definitions"]
|
|
assert forge.easy_brief("Character 2 waves", cast) == "<Subject 2> waves"
|
|
|
|
|
|
def test_separate_characters_beyond_four_stay_separate():
|
|
cast = forge.easy_cast([pic(f"character-{n}") for n in range(1, 10)])
|
|
assert len(cast["subjects"]) == 9
|
|
assert cast["subjects"][-1]["name"] == "Character 9"
|
|
assert forge.easy_brief("Character 9 waves", cast) == "<Subject 9> waves"
|
|
|
|
|
|
def test_unknown_label_is_character_1():
|
|
assert forge.easy_cast([pic("dragon")])["subjects"][0]["name"] == "Character 1"
|
|
|
|
|
|
def test_brief_names_become_tags():
|
|
cast = forge.easy_cast(REFS)
|
|
assert forge.easy_brief("Character 1 pours tea for character 2 in the place while Character 3 waits", cast) == \
|
|
"<Subject 1> pours tea for <Subject 2> in <Subject 3> while Character 3 waits"
|
|
assert forge.easy_brief("It takes place at night, in place of the party.", cast) == "It takes place at night, in place of the party."
|
|
assert forge.easy_brief("outfit from picture 4, not image 9, and <Picture 2>", cast) == "outfit from <Picture 4>, not image 9, and <Picture 2>"
|
|
|
|
|
|
def test_user_message_has_cast_not_picture_lines_and_no_attached_claim():
|
|
message = forge.build_user_message(forge.load_bundle(), "Character 1 waves", "REF2VA", 10, 5, "balanced",
|
|
REFS + [{"kind": "audio", "duration_seconds": 4}], False, cast=forge.easy_cast(REFS))
|
|
assert 'Brief: "<Subject 1> waves"' in message
|
|
assert '<Subject 2> is "Character 2" in the brief, a character, shown in <Picture 1> and <Picture 4>' in message
|
|
assert "keep: tired" in message
|
|
assert "<Picture 5> is the first frame" in message
|
|
assert "<Audio 1>" in message
|
|
assert "- <Picture 1> ·" not in message
|
|
assert "is attached" not in message
|
|
assert "Director output canvas is unknown" in message
|
|
|
|
|
|
def test_pose_custom_and_instructions_in_easy_mode():
|
|
refs = [pic("character-1", instructions="the red coat matters"), pic("pose", instructions="arms crossed"),
|
|
pic("custom", role="custom", instructions="use only the lighting"),
|
|
{"kind": "image", "role": "subject", "instructions": "a saved reference", "saved_reference": True}]
|
|
cast = forge.easy_cast(refs)
|
|
assert [s["name"] for s in cast["subjects"]] == ["Character 1"]
|
|
assert [(u["picture"], u["which"]) for u in cast["uses"]] == [(2, "pose"), (3, "custom")]
|
|
message = forge.build_user_message(forge.load_bundle(), "Character 1 in the pose of picture 2", "REF2VA", 10, 5, "balanced",
|
|
refs, False, output_canvas={"width": 1280, "height": 720}, cast=cast)
|
|
assert 'Brief: "<Subject 1> in the pose of <Picture 2>"' in message
|
|
assert "instructions: the red coat matters" in message
|
|
assert "<Picture 2> gives a pose only" in message and "instructions: arms crossed" in message
|
|
assert "<Picture 3> is used only as its instructions say · instructions: use only the lighting" in message
|
|
# The saved reference has no label, so it keeps its own reference line.
|
|
assert "- <Picture 4> · subject" in message and "a saved reference" in message
|
|
assert "aspect ratio 16:9" in message
|
|
segments = {"Detailed description": "[Shot 1] <Subject 1> stands."}
|
|
forge.easy_segments(cast, segments)
|
|
assert "<Picture 2> gives the pose only" in segments["Subject definitions"]
|
|
assert "not identity, clothing or background. Arms crossed." in segments["Subject definitions"]
|
|
assert "<Picture 3>: Use only the lighting." in segments["Subject definitions"]
|
|
|
|
|
|
def test_shots_line_and_cut_repair():
|
|
assert forge.shots_line("Auto") is None and forge.shots_line(None) is None
|
|
assert "exactly 3 shots" in forge.shots_line("3")
|
|
segments = {"Detailed description": "[Shot 1] A.\n[Shot 2] At 00:04.000, B.\n[Shot 3] At 00:10.000, C."}
|
|
assert forge.repair_cut_times(10, segments) == [("00:10.000", "00:07.000")]
|
|
assert "[Shot 3] At 00:07.000" in segments["Detailed description"]
|
|
assert forge.shot_count_warning("2", segments) and forge.shot_count_warning("3", segments) is None
|
|
|
|
|
|
def test_shot_boxes_join_the_idea():
|
|
assert forge.fold_shot_briefs("Rain.", "2", ["She runs.", " He turns. "]) == "Rain.\nShot 1: She runs.\nShot 2: He turns."
|
|
assert forge.fold_shot_briefs("x", "3", ["a", "", "c"]) == "x\nShot 1: a\nShot 3: c"
|
|
assert forge.fold_shot_briefs("x", "1", ["a", "b"]) == "x\nShot 1: a"
|
|
assert forge.fold_shot_briefs("x", "Auto", ["a"]) == "x"
|
|
assert forge.fold_shot_briefs("", "2", ["a", "b"]) == "Shot 1: a\nShot 2: b"
|
|
assert forge.fold_shot_briefs(" x ", None, None) == "x"
|
|
|
|
|
|
def test_segments_written_in_code():
|
|
cast = forge.easy_cast(REFS)
|
|
segments = {
|
|
"Subject definitions": "<Subject 1>: tired, slow blinks\n<Subject 2> — stiff and formal",
|
|
"Summary": "S.",
|
|
"Detailed description": "integrated_multimodal_description: Style.\n\n[Shot 1] <Subject 3>, <Subject 1> sits.\n\n[Shot 2] At 00:03.000, <Subject 1> pours.",
|
|
"Soundscape": "Rain.", "Music": "Koto.",
|
|
}
|
|
warnings = forge.easy_segments(cast, segments)
|
|
defs, ret = segments["Subject definitions"], segments["Retention analysis"]
|
|
assert defs.startswith("<Subject 1> is the character in <Picture 3>; keep their appearance exactly as the pictures show. In this scene: tired, slow blinks.")
|
|
assert "<Subject 2> is the character in <Picture 1> and <Picture 4>;" in defs
|
|
assert "<Subject 3> is the place in <Picture 2>, where the video happens" in defs
|
|
assert "<Picture 5> is the first frame of [Shot 1]." in defs
|
|
assert "<Subject 1> (appears in [Shot 1]-[Shot 2])" in ret
|
|
assert "<Subject 3> (appears in [Shot 1]-[Shot 2])" in ret
|
|
assert "tired" not in ret
|
|
assert len(warnings) == 1 and "Character 2" in warnings[0]
|
|
fields = forge.builder_fields(segments, "REF2VA")
|
|
assert fields["ref"]["retention_analysis"] == ret
|
|
|
|
|
|
def test_bundle_has_easy_mode_without_retention():
|
|
bundle = forge.load_bundle()
|
|
assert "Retention analysis" not in bundle["modes"][forge.EASY_MODE]["segments"]
|
|
assert "Subject definitions" in bundle["modes"][forge.EASY_MODE]["segments"]
|
|
|
|
|
|
def test_style_line_never_guesses_a_medium():
|
|
shots = "[Shot 1] <Subject 1> reads."
|
|
guessed = "Live-action, cinematic, warm interior light.\n\n" + shots
|
|
assert forge.keep_reference_look(guessed, "Character 1 reads") == "Keeps the look of the reference pictures.\n\n" + shots
|
|
assert forge.keep_reference_look(guessed, "live-action, cinematic: Character 1 reads") == guessed
|
|
fine = "Keeps the look of the reference pictures, warm indoor light.\n\n" + shots
|
|
assert forge.keep_reference_look(fine, "Character 1 reads") == fine
|
|
assert forge.keep_reference_look("[Shot 1] An animated wave.", "x") == "[Shot 1] An animated wave."
|
|
|
|
|
|
def test_runaway_is_refused_not_applied():
|
|
words = " ".join(["absurd ridiculous nonsensical illogical unreasonable"] * 40)
|
|
good = {"Summary": "A short summary.", "Detailed description": "[Shot 1] She reads. The camera holds."}
|
|
assert forge.runaway(good) is None
|
|
assert forge.runaway({**good, "Soundscape": "Room tone. " + words})[0] == "Soundscape"
|
|
raw = "===SEGMENT: Summary===\nS.\n===SEGMENT: Detailed description===\n[Shot 1] " + words
|
|
try:
|
|
forge.parse_segments(raw, ["Summary", "Detailed description", "Soundscape", "Music"])
|
|
assert False, "expected a runaway error"
|
|
except forge.ForgeError as exc:
|
|
assert exc.code == "runaway" and "Soundscape, Music" in exc.message
|
|
|
|
|
|
def test_music_warning_preserves_explicit_and_multilingual_requests():
|
|
bundle = forge.load_bundle()
|
|
for brief in ("She reads on the bed.", "She reads with a flute accompaniment.", "Eine Frau liest, mit Hintergrundmusik."):
|
|
segments = {"Music": "A soft flute accompaniment."}
|
|
assert forge.music_request_warning(bundle, brief, segments)
|
|
assert segments["Music"] == "A soft flute accompaniment."
|
|
assert not forge.music_request_warning(bundle, "Soft piano music plays.", {"Music": "Soft piano notes."})
|
|
assert not forge.music_request_warning({"music_words": None}, "She reads.", {"Music": "Soft piano notes."})
|
|
assert not forge.music_request_warning(bundle, "It takes place at night.", {"Music": "N/A"})
|
|
|
|
|
|
def test_runaway_allows_quoted_sentences_and_dialogue_tags():
|
|
sentence = 'A placard reads "Please use the main entrance while repairs to this side doorway are in progress."'
|
|
for body in ("[Shot 1] " + " ".join([sentence] * 6), "[Shot 1] " + " ".join(['<d>"Please use the main entrance while repairs to this side doorway are in progress."</d>'] * 6)):
|
|
assert len(body) > 500
|
|
assert forge.runaway({"Detailed description": body}) is None
|
|
|
|
|
|
def test_cut_repair_keeps_millisecond_order_near_clip_end():
|
|
for stamps in (("09.800", "10.000", "10.000"), ("09.950", "10.000")):
|
|
segments = {"Detailed description": "[Shot 1] A. " + " ".join(f"[Shot {n}] At 00:{stamp}, B." for n, stamp in enumerate(stamps, 2))}
|
|
assert forge.repair_cut_times(10, segments)
|
|
fields = forge.builder_fields(segments, "REF2VA")
|
|
assert forge.check_prompt(fields, "REF2VA", 10, "", 7000) == []
|
|
cuts = [int(m.group(2)) * 60 + float(m.group(3)) for m in forge._CUT.finditer(segments["Detailed description"])]
|
|
assert all(a < b < 10 for a, b in zip([0, *cuts], cuts))
|
|
segments = {"Detailed description": "[Shot 1] A. [Shot 2] At 00:09.999, B. [Shot 3] At 00:10.000, C. [Shot 4] At 00:10.000, D."}
|
|
forge.repair_cut_times(10, segments)
|
|
assert forge.check_prompt(forge.builder_fields(segments, "REF2VA"), "REF2VA", 10, "", 7000)
|
|
|
|
|
|
def test_typed_ref2va_validation_uses_the_typed_sections():
|
|
from helper_minimax_h3_prompt_builder import validate_builder_state
|
|
state = {"mode": "REF2VA", "prompt_mode": "simple", "ref": {}, "simple_prompt": "A woman waves."}
|
|
assert validate_builder_state(state) == []
|
|
state["simple_prompt"] = "subject_definitions:\n<Subject 1> is a woman.\nsummary:\nA short scene.\nretention_analysis:\nN/A\ndetailed_description:\n[Shot 1] She waves.\noverall_soundscape:\nRoom tone.\nnon_diegetic_music:\nN/A"
|
|
assert validate_builder_state(state) == []
|
|
state["simple_prompt"] = state["simple_prompt"].replace("A short scene.", "")
|
|
assert validate_builder_state(state) == [{"level": "warn", "msg": "REF2VA summary is empty."}]
|
|
state["prompt_mode"] = "structured"
|
|
assert len(validate_builder_state(state)) == 2
|
|
|
|
|
|
def test_mixed_and_saved_references_keep_full_definitions(monkeypatch):
|
|
class Backend:
|
|
base = "https://fixture.invalid"
|
|
def can_see(self, name):
|
|
return False
|
|
def chat(self, *args):
|
|
segments = {
|
|
"Subject definitions": "<Subject 1> from <Picture 1>. <Subject 2> from <Picture 2>. <Audio 1> supplies the voice. <Video 1> supplies motion.",
|
|
"Summary": "A ten-second scene.",
|
|
"Retention analysis": "<Audio 1>: fully_copy. <Video 1>: reference. <Subject 2>: fully_preserved.",
|
|
"Detailed description": "[Shot 1] <Subject 1> waves, following <Video 1>, while <Audio 1> speaks.",
|
|
"Soundscape": "Room tone.", "Music": "N/A",
|
|
}
|
|
return "\n".join(f"===SEGMENT: {label}===\n{body}" for label, body in segments.items()), {}
|
|
def unload(self, name):
|
|
return True
|
|
monkeypatch.setattr(forge, "backends", lambda settings: {"openai": Backend()})
|
|
for extra in ([{"kind": "audio"}, {"kind": "video"}], [{"kind": "image", "saved_reference": True}]):
|
|
result = forge._generate({"mode": "REF2VA", "model": "openai:fixture", "brief": "She waves.", "duration": 10, "easy": True, "references": [pic("character-1"), *extra]}, None, None, None)
|
|
assert result["easy"] is False
|
|
assert "<Audio 1> supplies the voice" in result["fields"]["ref"]["subject_definitions"]
|
|
assert "<Video 1>: reference" in result["fields"]["ref"]["retention_analysis"]
|
|
assert "<Subject 2> from <Picture 2>" in result["fields"]["ref"]["subject_definitions"]
|