Free, open-source AI video generator for Mac. Run LTX-2.5 and MiniMax H3 (Hailuo) locally on Apple Silicon: text-to-video and image-to-video with sound, lip-sync, character LoRA training, songs from lyrics (YuE2) and image editing (Qwen-Image-Edit). One-click install via Pinokio. Offline: no cloud, no API key, no NVIDIA GPU.
Python ai-filmmakingai-videoai-video-generatorapple-siliconhailuoimage-to-videolip-synclocal-ailora-trainingltx-2ltx-2-5ltx-videomacosminimax-h3mlxmusic-generationoffline-aipinokiotext-to-videovideo-generation
release(v4.12.1): a One Shot with no beats is the prompt in every window; the hold names what continues
mrbizarrob322e18fdf6a
diff --git a/README.md b/README.mdindex 4f37d02..8d3f9ac 100644--- a/README.md+++ b/README.md@@ -19,7 +19,7 @@</p>-> **Current release: v4.12.0.** A One Shot with a talking character now keeps the voice on the mouth across the joins: each part hands off on a frame where the character is still mid-word, the last word carries over the next part, every spoken part is measured for lip-sync and retaken up to twice, and the joined clip no longer drifts its sound ahead of the picture. Dialogue between two characters has a recipe that works (one face on screen at a time, a shot that carries a line frames only the speaker), written up in `docs/LTX_DIALOGUE.md`. Includes 4.11.x. Full notes on the [releases page](https://github.com/mrbizarro/Phosphene/releases).+> **Current release: v4.12.1.** A One Shot with no beats written is a One Shot of the prompt, sustained for the whole take (it used to render the hold sentence, a clip about nothing). Includes 4.12.0: a One Shot with a talking character keeps the voice on the mouth across the joins, every spoken part is measured and retaken, the joined clip no longer drifts its sound, and two characters can talk (`docs/LTX_DIALOGUE.md`). Full notes on the [releases page](https://github.com/mrbizarro/Phosphene/releases).## Overviewdiff --git a/VERSION b/VERSIONindex 815588e..53cf85e 100644--- a/VERSION+++ b/VERSION@@ -1 +1 @@-4.12.0+4.12.1diff --git a/docs/API.md b/docs/API.mdindex df5fc3a..65c9e39 100644--- a/docs/API.md+++ b/docs/API.md@@ -142,7 +142,9 @@ Submit multiple prompts as one batch.A clip longer than a single pass, on either engine. `take_seconds` is one of30 / 45 / 60 / 90 / 120; `beats` is a JSON list (or newline text) with oneprompt per five seconds — a blank beat holds the previous moment, extras are-dropped. The take overrules the engine's own length field. On BOTH engines it+dropped. **No beats at all** (the field absent, empty, or only blank lines) means the+`prompt` is every beat: the shot is the prompt sustained for the whole take. Before+4.12.1 that case rendered every window as the hold sentence — a take about nothing. The take overrules the engine's own length field. On BOTH engines itrenders as PARTS chained by last-frame handoff — `run_take_job_inner` runs oneordinary render per part, each starting from the last frame of the one before(i2v, anchored), hides the parts from the gallery and joins them intodiff --git a/docs/PROMPTING.md b/docs/PROMPTING.mdindex 0e57c68..e8e66d0 100644--- a/docs/PROMPTING.md+++ b/docs/PROMPTING.md@@ -4,6 +4,7 @@ Copy this whole page into your own assistant (ChatGPT, Claude, Codex) and ask it## One Shot (a clip longer than one pass)+- **You do not have to write beats.** Leave the box empty and the prompt carries the whole shot, sustained for every five seconds. Write beats when you want the shot to *go* somewhere: each line directs its own five seconds, a blank line holds the moment before it.- **One beat = five seconds = one thing that happens.** Lead with the movement: what the subject does, what the camera does, what enters the frame.- **A scene change gets a beat of its own.** Going through a door, into a room, onto a bridge: give it its own five seconds, and put the *approach* in the beat before it. A scene change hidden at the end of a busy beat is the one the model skips.- **A reveal gets a beat of its own.** "The camera comes round to his face" works when it is the whole beat, not the tail of a beat that also carries the table, the wife and the hat.diff --git a/docs/STATE.md b/docs/STATE.mdindex 491a190..32f61c4 100644--- a/docs/STATE.md+++ b/docs/STATE.md@@ -173,6 +173,12 @@> shitty on the faces" — and the two-character experiment (bizarrotrn +> eltrumpo at one table, Q8 Pro, camera lock) is the first render on this code.+> **🚀 2026-09-08 — v4.12.1 ships: a One Shot with no beats is the prompt.** A first-day+> field report: a long prompt, an empty beats box, and every window rendered the+> hold sentence ("Nothing new begins…") — a 30 s clip about nothing, on H3 too.+> `make_job` now makes the prompt every beat when none is written; `TAKE_HOLD`+> names what continues; the beats hint says when the prompt carries the shot.+> **🚀 2026-09-08 — v4.12.0 ships: the talking one shot.** `take_handoff=speech` as a> J-cut (picture cut on a talking frame, the last word mixed over the next part),> the lip-sync gate with two retakes that only judges parts meant to speak, thediff --git a/mlx_ltx_panel.py b/mlx_ltx_panel.pyindex 7189278..3100bbf 100644--- a/mlx_ltx_panel.py+++ b/mlx_ltx_panel.py@@ -15899,7 +15899,11 @@ _WEATHER_WORDS = ("rain", "drizzle", "storm", "snow", "fog", "mist", "overcast",TAKE_DRIFT_MAX = 0.28 # |Δ mean luma| between a part's first and last frame, 0..1# A blank beat means "hold": H3's runner refuses an empty window prompt, so# on H3 a blank beat is spelled out (LTX's windows chain holds on its own).-TAKE_HOLD = "Nothing new begins: the settled state of the previous moment continues."+# The old sentence was "Nothing new begins: the settled state of the previous+# moment continues." — and H3 rendered exactly that, a clip about nothing+# (field report, 2026-09-08: "this is what a video about 'nothing new' looks+# like"). A hold must name what continues, not what does not.+TAKE_HOLD = "The scene continues exactly as it was: the same subject, the same motion, the same light, carrying on."TAKE_DRIFT_SAMPLE_SEC = 0.25 # how far in from each end the frames are read@@ -20855,8 +20859,16 @@ def make_job(form: dict[str, list[str]] | dict[str, str], *,else:_lock_on = f("take_light_lock", "on").strip().lower() not in ("off", "0", "false", "no")_lock = take_light_lock(f("prompt", "")) if _lock_on else ""+ _beats_written = take_beats(f("beats", ""), _take["beats"])+ if not any(_beats_written):+ # No beat written: the PROMPT is the shot, in every window. Before+ # this a plain prompt with an empty beats box rendered every+ # window as the hold sentence — a 30 s take about "nothing new"+ # (field report, 2026-09-08). Blank beats still hold, but only+ # when the user wrote some beats and left the others blank.+ _beats_written = [f("prompt", "").strip()] * _take["beats"]_take["beat_prompts"] = [(b + " " + _lock) if (b and _lock) else b- for b in take_beats(f("beats", ""), _take["beats"])]+ for b in _beats_written]_take["light_lock"] = _lock_take["retake"] = f("take_retake", "on").strip().lower() not in ("off", "0", "false", "no")# The camera, once, for the whole shot (take_camera_lock).diff --git a/test_take.py b/test_take.pyindex b9afc93..3651689 100644--- a/test_take.py+++ b/test_take.py@@ -625,3 +625,22 @@ def test_lipsync_gate_only_judges_parts_that_speak(tmp_path, monkeypatch):p.run_take_job_inner(j)assert seen == [f"{j['id']}-p1", f"{j['id']}-p2", f"{j['id']}-p3"] # no -p1l: nothing was retakenassert j["take_lipsync"] == [None, None, None]+++def test_take_with_no_beats_is_the_prompt_in_every_window():+ """A prompt and an empty beats box is a take OF THE PROMPT, on both engines —+ not six windows of the hold sentence (field report, 2026-09-08)."""+ for beats in ("", json.dumps([]), json.dumps(["", "", ""]), "\n\n"):+ q = p.make_job({"mode": "t2v", "engine": "h3", "prompt": "A hen skates down Broadway at night.",+ "take_seconds": "30", "beats": beats})["params"]+ assert len(q["take"]["beat_prompts"]) == 6+ assert all(b.startswith("A hen skates down Broadway at night.") for b in q["take"]["beat_prompts"]), beats+ assert all(c.startswith("A hen skates") for c in q["h3_chain_prompts"]), beats+ assert not any(p.TAKE_HOLD in c for c in q["h3_chain_prompts"])+ q = p.make_job({"mode": "t2v", "engine": "ltx", "prompt": "A hen skates down Broadway at night.", "take_seconds": "30"})["params"]+ assert all(b.startswith("A hen skates") for b in q["take"]["beat_prompts"])+ # some beats written: the blanks still hold, as before+ q = p.make_job({"mode": "t2v", "engine": "h3", "prompt": "A hen skates.", "take_seconds": "30",+ "beats": json.dumps(["she pushes off", "", "a van sweeps past"])})["params"]+ assert q["h3_chain_prompts"][0].startswith("she pushes off") and q["h3_chain_prompts"][1].startswith(p.TAKE_HOLD)+ assert "nothing new" not in p.TAKE_HOLD.lower() # the hold names what continues, not what does notdiff --git a/webapp/js/characters.js b/webapp/js/characters.jsindex f5cb99f..03ac589 100644--- a/webapp/js/characters.js+++ b/webapp/js/characters.js@@ -2677,9 +2677,14 @@ function beatsInput() {if (hint && n) {const filled = lines.filter(Boolean).length;const extra = lines.length - n;- hint.textContent = `${n} lines of 5 s · ${filled} written`- + (filled < n ? ' · leave a line blank to hold on the scene' : '')- + (extra > 0 ? ` · ${extra} extra line${extra === 1 ? '' : 's'} will be dropped` : '');+ // Zero written is not an error: the prompt is the whole shot. Say so,+ // because a box that reads "0 written" next to a Generate button looks+ // like something the user forgot (field report, 2026-09-08).+ hint.textContent = filled === 0+ ? `${n} lines of 5 s · none written — the prompt carries the whole shot; write a line to direct a moment`+ : `${n} lines of 5 s · ${filled} written`+ + (filled < n ? ' · leave a line blank to hold on the scene' : '')+ + (extra > 0 ? ` · ${extra} extra line${extra === 1 ? '' : 's'} will be dropped` : '');}}let _takeRefreshSeq = 0;