feat: shot camera can reframe within a shot: between two shortcuts it was not cut at, the camera changes view and focus without a cut
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
digital-garbage
co-authored by
Claude Opus 5.5
Claude Opus 5.5
parent
e1dfdd80d9
commit
9cb0932f03
@@ -6,6 +6,7 @@ pick the move. Off by default:
|
||||
|
||||
* cut times: shots after the first open with their time
|
||||
* cut within the same shot: "[Shot 1] DANCE KISS" and "[Shot 1] DANCE DANCE" are cut between the shortcuts, so each part gets a time, a view and a focus. Its chance slider is hidden until this is on
|
||||
* reframe within the shot: at a boundary that was not cut, the camera changes view and focus with no cut, in one continuous take
|
||||
* views: your own trusted list (side view, from above, POV...), never on shot 1, never the same twice in a row
|
||||
* camera moves: one to three per shot, aimed at the most specific thing in it ("<Subject 1>'s hand", not "<Subject 1>")
|
||||
* liked details: a short detail you liked ("detailed lips") may be added, once, at the end of a later shot that already names that body part. Its chance slider is hidden until this is on
|
||||
@@ -41,6 +42,11 @@ SETTINGS = {
|
||||
"shot_cuts_chance": {"type": "float", "default": 0.5, "min": 0.0, "max": 1.0, "step": 0.05, "label": "Chance of a cut within the shot",
|
||||
"hint": "1 = always cut between those shortcuts, 0 = never. Ratings nudge the values in between.",
|
||||
"when": {"cut_same_shot": True}},
|
||||
"reframe_same_shot": {"type": "bool", "default": False, "label": "Reframe within the shot",
|
||||
"hint": "Where one shortcut ends and the next begins inside a shot that is not cut, the camera moves to a new view and focus without a cut: one continuous take, the actions follow on. A shot that reframes gets no extra camera move."},
|
||||
"reframe_chance": {"type": "float", "default": 0.5, "min": 0.0, "max": 1.0, "step": 0.05, "label": "Chance of a reframe",
|
||||
"hint": "1 = always reframe between those shortcuts, 0 = never. Ratings nudge the values in between.",
|
||||
"when": {"reframe_same_shot": True}},
|
||||
"shot_views": {"type": "bool", "default": False, "label": "Shot views",
|
||||
"hint": "Later shots open on a view (side, from above, POV…), never the same twice in a row."},
|
||||
"shot_views_chance": {"type": "float", "default": 0.4, "min": 0.0, "max": 1.0, "step": 0.05, "label": "Chance a shot gets a view",
|
||||
|
||||
@@ -717,6 +717,60 @@ def add_shot_cuts(text, seconds, seed=0, chance=0.5, pieces=()):
|
||||
return "".join(out), info
|
||||
|
||||
|
||||
# ── reframes ────────────────────────────────────────────────────────────────────────
|
||||
# A shortcut boundary inside a shot that is NOT cut can turn the camera instead: one continuous
|
||||
# take, the actions follow on, and only the view and the focus change. Run after the cuts, so
|
||||
# the boundaries left are exactly the ones that were kept whole.
|
||||
REFRAMES = ("Without a cut, the camera smoothly moves to {view}{x}.",
|
||||
"Without a cut, the camera glides around to {view}{x}.",
|
||||
"Without a cut, the camera drifts over to {view}{x}.")
|
||||
_REFRAME = re.compile(r"\s*Without a cut, the camera [^.!?]*[.!?]", re.I)
|
||||
|
||||
|
||||
def _a_view(view):
|
||||
return "a " + (view if view.startswith("POV") else view.lower())
|
||||
|
||||
|
||||
def add_reframes(text, seed=0, chance=0.5, pieces=(), stats=None):
|
||||
"""-> (new prompt, info). At each shortcut boundary inside a shot, with odds `chance`, a
|
||||
sentence moves the camera to a new view (never the one just used) aimed at the most specific
|
||||
thing in what follows, without a cut. info = {"added": [{"shot", "view", "traits",
|
||||
"target"}], "arms": ["reframe:yes"|"reframe:no", ...]}."""
|
||||
marks = list(SHOT.finditer(text or ""))
|
||||
info = {"added": [], "arms": []}
|
||||
if not marks:
|
||||
return text, info
|
||||
out, last = [text[:marks[0].start()]], None
|
||||
for i, m in enumerate(marks):
|
||||
end = marks[i + 1].start() if i + 1 < len(marks) else len(text)
|
||||
picture, sound = _split_sound(text[m.end():end])
|
||||
points = _cut_points(picture, pieces)
|
||||
chosen = []
|
||||
for n, at in enumerate(points):
|
||||
rng = random.Random(f"{seed}:reframe:{i}:{n}")
|
||||
take = rng.random() < chance
|
||||
info["arms"].append("reframe:yes" if take else "reframe:no")
|
||||
if not take:
|
||||
continue
|
||||
following = picture[at:(points[n + 1] if n + 1 < len(points) else len(picture))]
|
||||
traits = view_traits(following)
|
||||
allowed = allowed_views(traits)
|
||||
pool = [v for v in allowed if v != last] or allowed
|
||||
view = rng.choices(pool, weights=[_view_weight(v, traits, stats) for v in pool])[0]
|
||||
cands = [c for c in candidates(own_words_removed(following)) if _score(c) > 0.5]
|
||||
target = _name(_draw(cands, None, None, 1)[0]) if cands else None
|
||||
sentence = rng.choice(REFRAMES).format(view=_a_view(view), x=f", framing {target}" if target else "")
|
||||
chosen.append((at, sentence))
|
||||
info["added"].append({"shot": int(m.group(1)), "view": view, "traits": traits, "target": target})
|
||||
last = view
|
||||
for at, sentence in reversed(chosen): # right to left: earlier indexes stay valid
|
||||
head = picture[:at].rstrip()
|
||||
picture = head + ("" if head.endswith((".", "!", "?")) else ".") + " " + sentence + " " \
|
||||
+ _fresh_piece(picture[at:])
|
||||
out.append(m.group(0) + picture + sound)
|
||||
return "".join(out), info
|
||||
|
||||
|
||||
# ── views ───────────────────────────────────────────────────────────────────────────
|
||||
# Chosen, not detected: a prompt seldom says which view it wants. The user's own trusted
|
||||
# wording, verbatim. Shot 1 is skipped: it may sit on a reference image or a pinned first
|
||||
@@ -799,7 +853,7 @@ def view_options(text, stats=None):
|
||||
traits = view_traits(picture)
|
||||
cands = sorted(((v, _view_weight(v, traits, stats)) for v in allowed_views(traits)),
|
||||
key=lambda c: -c[1])
|
||||
stated = VIEW_STATED.search(CUT.sub("", picture))
|
||||
stated = VIEW_STATED.search(CUT.sub("", _REFRAME.sub("", picture)))
|
||||
out.append({"shot": int(m.group(1)), "key": key, "traits": traits,
|
||||
"already": bool(stated), "stated": stated.group(0) if stated else "",
|
||||
"candidates": [{"view": v, "score": round(w, 2)} for v, w in cands],
|
||||
@@ -829,7 +883,7 @@ def add_shot_views(text, seed=0, chance=0.4, stats=None, choices=None, skipped=N
|
||||
picture, sound = _split_sound(body)
|
||||
key = keys(picture)
|
||||
view = None
|
||||
stated = VIEW_STATED.search(CUT.sub("", picture)) if (i or lone) else None
|
||||
stated = VIEW_STATED.search(CUT.sub("", _REFRAME.sub("", picture))) if (i or lone) else None
|
||||
if stated and skipped is not None:
|
||||
skipped.append(f"shot {m.group(1)}: already states “{stated.group(0)}”")
|
||||
if (i or lone) and not stated:
|
||||
@@ -872,7 +926,7 @@ def own_words_removed(picture):
|
||||
become a target or a topic."""
|
||||
picture = _OUR_OPENER.sub("", picture, count=1)
|
||||
picture = _CUT_WORDS.sub("", picture, count=1)
|
||||
return _VIEW_SENTENCE.sub("", picture, count=1)
|
||||
return _REFRAME.sub("", _VIEW_SENTENCE.sub("", picture, count=1))
|
||||
|
||||
|
||||
_OUR_OPENER = re.compile(r"^\s*At \d\d:\d\d\.\d{3},\s*(?:" + "|".join(re.escape(o) for o in CUT_OPENERS)
|
||||
@@ -888,6 +942,7 @@ def content_fingerprint(text):
|
||||
body = re.sub(r"\[\s*Shot\s+\d+\s*\]", " ", text or "", flags=re.I)
|
||||
body = re.sub(r"At \d\d:\d\d\.\d{3},\s*(?:" + "|".join(re.escape(o) for o in CUT_OPENERS) + r")\.",
|
||||
" ", body, flags=re.I)
|
||||
body = _REFRAME.sub(" ", body)
|
||||
body = re.sub(r"\s+", " ", body)
|
||||
body = re.sub(r"(?:^|(?<=[.!?]\s))(?:" + "|".join(re.escape(v) for v in VIEWS) + r")\.", " ", body)
|
||||
body = re.sub(r"\s+", " ", body).strip().lower()
|
||||
|
||||
@@ -171,7 +171,8 @@ def learn_views(decisions):
|
||||
def arm_stats():
|
||||
"""{arm: (good, bad)} for what ratings taught about the camera's own choices: "move:yes" /
|
||||
"move:no" (a shot got a move / was left alone), "style:hold|travel|k1|k2|k3", "word:<lemma>",
|
||||
"split:yes|no" (a shortcut boundary inside a shot was cut / kept whole)."""
|
||||
"split:yes|no" (a shortcut boundary inside a shot was cut / kept whole), "reframe:yes|no" (a kept
|
||||
boundary turned the camera without a cut / did not)."""
|
||||
return {k: (float(v[0]), float(v[1])) for k, v in (_read().get("arms") or {}).items()
|
||||
if isinstance(v, (list, tuple)) and len(v) == 2}
|
||||
|
||||
@@ -341,6 +342,13 @@ def split_chance(chance):
|
||||
return max(0.0, min(1.0, chance * _lean("split:yes", "split:no")))
|
||||
|
||||
|
||||
def reframe_chance(chance):
|
||||
"""The chance of reframing at a kept boundary, tilted by whether reframed runs were liked (0 and 1 kept as set)."""
|
||||
if chance <= 0.0 or chance >= 1.0:
|
||||
return max(0.0, min(1.0, chance))
|
||||
return max(0.0, min(1.0, chance * _lean("reframe:yes", "reframe:no")))
|
||||
|
||||
|
||||
def _save(data):
|
||||
path = _path()
|
||||
os.makedirs(os.path.dirname(path), exist_ok=True)
|
||||
|
||||
@@ -52,6 +52,13 @@ def rewrite(text, values, seconds=None, pieces=None):
|
||||
text, info = engine.add_shot_cuts(text, seconds, seed=seed, chance=chance, pieces=_pieces() if pieces is None else pieces)
|
||||
arms += info.get("arms", [])
|
||||
said.append(f"cuts at {', '.join(info['times'])}" + (f" ({info['before']}→{info['after']} shots)" if info["after"] != info["before"] else "") if info["times"] else f"cuts: {info['why']}")
|
||||
if values.get("reframe_same_shot") and engine.SHOT.search(text or ""):
|
||||
text, info = engine.add_reframes(text, seed=seed, chance=memory.reframe_chance(float(values.get("reframe_chance", 0.5))),
|
||||
pieces=_pieces() if pieces is None else pieces, stats=memory.view_stats())
|
||||
arms += info["arms"]
|
||||
views += [{"view": a["view"], "traits": a["traits"]} for a in info["added"]]
|
||||
said.append("reframes: " + (", ".join(f"shot {a['shot']} {a['view']}" + (f" on {a['target']}" if a["target"] else "") for a in info["added"])
|
||||
if info["added"] else "none" + ("" if info["arms"] else " (no shortcut boundary inside a shot)")))
|
||||
if values.get("shot_views") and engine.SHOT.search(text or ""):
|
||||
skipped = []
|
||||
text, added = engine.add_shot_views(text, seed=seed, chance=float(values.get("shot_views_chance", 0.4)), stats=memory.view_stats(), skipped=skipped)
|
||||
@@ -105,7 +112,7 @@ class FunPackShotCamera(io.ComfyNode):
|
||||
@classmethod
|
||||
def execute(cls, text: str, length: int, frame_rate: float, settings=None) -> io.NodeOutput:
|
||||
values = _values(settings)
|
||||
if not any(values.get(k) for k in ("camera_moves", "shot_cuts", "shot_views", "cut_same_shot", "detail_notes")) or not engine.SHOT.search(text or ""):
|
||||
if not any(values.get(k) for k in ("camera_moves", "shot_cuts", "shot_views", "cut_same_shot", "reframe_same_shot", "detail_notes")) or not engine.SHOT.search(text or ""):
|
||||
return io.NodeOutput(text, "unchanged")
|
||||
try:
|
||||
out, said, chose = rewrite(text, values, seconds=length / frame_rate if frame_rate else None)
|
||||
|
||||
@@ -0,0 +1,69 @@
|
||||
"""Reframe within the shot: a kept shortcut boundary turns the camera in one continuous take, never a cut."""
|
||||
|
||||
from .. import engine, memory, nodes
|
||||
|
||||
DANCE = "She sways her hips from side to side, smiling at the room."
|
||||
WAVE = "She raises her hand and waves at the camera with her fingers."
|
||||
|
||||
|
||||
def test_reframe_turns_the_camera_between_shortcuts_without_a_cut(tmp_path, monkeypatch):
|
||||
monkeypatch.setenv("SHOT_CAMERA_MEMORY", str(tmp_path / "m.json"))
|
||||
text = f"[Shot 1] {DANCE} {WAVE}"
|
||||
out, said, chose = nodes.rewrite(text, {"reframe_same_shot": True, "reframe_chance": 1}, seconds=10, pieces=[DANCE, WAVE])
|
||||
assert out.count("[Shot ") == 1 and "cuts to" not in out, "one take: no new shot, no cut words"
|
||||
head, tail = out.split("Without a cut, the camera ")
|
||||
assert DANCE in head and WAVE in tail, "the reframe sits between the two actions, in order"
|
||||
assert "framing" in tail.split(".")[0], "aimed at what the next action is about"
|
||||
assert "reframes: shot 1" in said
|
||||
assert chose["arms"] == ["reframe:yes"] and chose["views"], "learned like the other camera choices"
|
||||
|
||||
|
||||
def test_the_view_never_repeats_and_respects_what_the_part_shows():
|
||||
text = f"[Shot 1] {DANCE} {DANCE} {DANCE} {DANCE}"
|
||||
out, info = engine.add_reframes(text, seed="s", chance=1, pieces=[DANCE])
|
||||
views = [a["view"] for a in info["added"]]
|
||||
assert len(views) == 3 and all(a != b for a, b in zip(views, views[1:]))
|
||||
assert all(v != "View from behind" for v in views), "a smiling face is not shown from behind"
|
||||
|
||||
|
||||
def test_zero_chance_adds_nothing_and_no_boundary_says_so(tmp_path, monkeypatch):
|
||||
monkeypatch.setenv("SHOT_CAMERA_MEMORY", str(tmp_path / "m.json"))
|
||||
text = f"[Shot 1] {DANCE} {WAVE}"
|
||||
assert engine.add_reframes(text, seed=1, chance=0, pieces=[DANCE, WAVE])[0] == text
|
||||
out, said, _ = nodes.rewrite("[Shot 1] A woman walks.", {"reframe_same_shot": True, "reframe_chance": 1}, seconds=5, pieces=[])
|
||||
assert out == "[Shot 1] A woman walks." and "no shortcut boundary inside a shot" in said
|
||||
|
||||
|
||||
def test_a_cut_boundary_is_not_also_reframed(tmp_path, monkeypatch):
|
||||
monkeypatch.setenv("SHOT_CAMERA_MEMORY", str(tmp_path / "m.json"))
|
||||
text = f"[Shot 1] {DANCE} {WAVE}"
|
||||
out, _, _ = nodes.rewrite(text, {"cut_same_shot": True, "shot_cuts_chance": 1, "reframe_same_shot": True, "reframe_chance": 1},
|
||||
seconds=10, pieces=[DANCE, WAVE])
|
||||
assert out.count("[Shot ") == 2 and "Without a cut" not in out
|
||||
|
||||
|
||||
def test_the_reframe_is_not_content_and_does_not_block_the_opening_view():
|
||||
text = f"[Shot 1] A woman walks in. [Shot 2] {DANCE} {WAVE}"
|
||||
out, _ = engine.add_reframes(text, seed=2, chance=1, pieces=[DANCE, WAVE])
|
||||
assert engine.content_fingerprint(out) == engine.content_fingerprint(text), "same prompt, same seed next run"
|
||||
shot2 = engine.shot_texts(out)[2]
|
||||
assert "Without a cut" not in shot2
|
||||
_, added = engine.add_shot_views(out, seed=0, chance=1)
|
||||
assert [a["shot"] for a in added] == [2], "a mid-shot reframe is not the view the shot opens on"
|
||||
|
||||
|
||||
def test_ratings_tilt_the_reframe_chance(tmp_path, monkeypatch):
|
||||
monkeypatch.setenv("SHOT_CAMERA_MEMORY", str(tmp_path / "m.json"))
|
||||
assert memory.reframe_chance(0.5) == 0.5
|
||||
memory.record_run("r", {"views": [], "arms": ["reframe:yes"]})
|
||||
memory.on_rating("r", "liked")
|
||||
assert memory.reframe_chance(0.5) > 0.5
|
||||
assert memory.reframe_chance(1.0) == 1.0, "always stays always"
|
||||
|
||||
|
||||
def test_cut_within_the_shot_at_zero_leaves_every_boundary_to_reframe(tmp_path, monkeypatch):
|
||||
monkeypatch.setenv("SHOT_CAMERA_MEMORY", str(tmp_path / "m.json"))
|
||||
text = f"[Shot 1] {DANCE} {WAVE} {DANCE}"
|
||||
out, _, _ = nodes.rewrite(text, {"cut_same_shot": True, "shot_cuts_chance": 0, "reframe_same_shot": True, "reframe_chance": 1},
|
||||
seconds=10, pieces=[DANCE, WAVE])
|
||||
assert out.count("[Shot ") == 1 and out.count("Without a cut") == 2
|
||||
Reference in New Issue
Block a user