One prompt → one song → a two-scene music video (YuE2 × MiniMax H3)

Score to Screen v7.0 Simple — changes since v6.2v7.0 is a new way to work: the whole MV is made from one screen. The 🎞 Timeline gets a left column with four tabs (🎬 Main, 🎵 Song, 🎚 Sound, ⚙ Settings), can fill the whole window, and opens by itself. New in it: 🎭 a cast of up to four people with their pictures, picked per scene; 🎵 the song in its own field; 🎚 the finished MV's sound (EQ, reverb, width, fades) and 🔁 remaking only the sound; ⚙ models, LoRAs and VAEs in one place. It also brings the hand cues made after v6.2 (🎸 the fretting hand timed, 🎻 bow and fingers, 🎥 a close-up of the fingers) and ⚡ one Turbo steps setting. Everything from v6.2 (🔒 Audio Lock, 🥁 adaptive drums, 🔁 remake one scene, 🥁 drum timing from the picture) stays as it was.🎞 One screen for the whole MV (v7.0)Why<ul><li>Every new feature added another node somewhere on the canvas. Making one MV meant finding the scene columns, the 🧑 / ⛓ group, the 🎛 switches, 🔧 Details and 📦 Models in turn, and a setting in the wrong node (an audio VAE in the video VAE loader) only showed up as an error halfway through the run</li></ul>What it does<ul><li>A left column with four tabs next to the track and the scene's text:<ul><li>🎬 Main: 🎭 the cast (below)</li><li>🎵 Song: the song's description, 🎤 does the person on screen sing, 🎵 a pre-made song and where it starts, 🎵 song only, 📦 the YuE2 model</li><li>🎚 Sound: the finished MV's sound (below)</li><li>⚙ Settings: ⚡ Turbo on / off and its steps, ⚡ the Turbo LoRA and its strength, 📦 the H3 model, ➕ up to three more LoRAs, 📝 the text encoder, 🎞 the video VAE, 🔊 the audio VAE, and 🎸🎻 playing &amp; camera (fretting hand timed, hand close-up at a solo, bow &amp; fingers, the drum cue style, 🔒 Audio Lock)</li></ul></li><li>⛶ Full screen: one button fills the window with the timeline, with its own ▶ Run. "🗗 Back down" or Esc brings back the bottom panel. The left column's width can be dragged</li><li>Opens by itself whenever a Score to Screen workflow is opened (ComfyUI starting, a file, a tab). ✕ closes it until the next workflow</li><li>▶ Next run, one line under the header: every scene / one scene again / 🎚 sound only / song only, the scenes and seconds, Turbo and steps, the cast, 🔒 and 🥁. Anything that would make the run stop or go wrong comes first, in orange: a VAE of the wrong kind, 🔁 with a run but no scene picked, 🎚 sound only without a run. The panel's ▶ Run refuses to start such a run; ComfyUI's own Run shows the same warning</li><li>The VAE lists only offer their own kind (audio VAEs for the audio VAE, video VAEs for the video VAE, judged by the file name). When none of a kind is found, or the node holds the other kind, the list says so instead of offering it</li><li>The panel is still only a second view of the workflow's own nodes: everything can also be set on the canvas as before</li></ul>🎭 The cast (v7.0)Why<ul><li>Until now there was one 🧑 character picture for every scene, and who plays what came only from the LLM's reading of each scene's text and picture. In a band MV the drummer could turn into a second guitarist from one scene to the next, and a second person had no picture to keep their face</li></ul>What it does<ul><li>🧑 A main character (S1, in every scene) and 👤 three more (S2–S4), each with a picture, a name, a part (sings, guitar, guitar + sings, bass, drums, keyboard, piano, violin, cello, dances ...) and a note for the LLM in any language</li><li>Per scene: a sub is only in the scenes you pick (1–4 on their card, or the chips under the selected scene's text: "🎭 In scene 2: ✓ S2 Ken S3 Mio"). The faces of who is in a scene show on its clip</li><li>To H3: the pictures of the people in the scene go in as reference pictures. &lt;Picture 1&gt; is the scene's own picture, &lt;Picture 2&gt; the main character, then the subs in that scene in S order; someone without a picture is skipped and the numbers close up (H3 does the same). The panel shows each person's number ("S1・Aya・🎻 violin・🖼 2", or "no picture")</li><li>In the prompt: one line per person, "S2 is Ken, the person shown in &lt;Picture 3&gt;, who plays the drums." (replacing the LLM's own line for that S number, or added after the other definitions), and "&lt;Picture 3&gt; shows S2 (Ken): S2 keeps exactly the same face, hairstyle and clothes as in &lt;Picture 3&gt; for the whole video."</li><li>To the cues: the part is written the way 🎼 ABC Timing reads it, so a drummer gets the drum cues, a singer the lip sync, a bassist the root notes</li><li>To the LLM: "(Who is in this scene - use exactly these S numbers and names in the definitions: S1 = Aya (the main character), plays the guitar and sings; S2 = Ken, plays the drums (blond).)" is added to what it reads (never to the lyrics-language check)</li><li>Only a picture for the main character (no name, no part): the LLM's own S1 line stays, as in v6</li><li>The 🧑 character picture moved from the ⛓ Scene starts node to the cast's main character</li></ul>Note: every extra picture adds reference tokens, so H3 gets a little slower with each person.🎵 The song has its own field (v7.0)<ul><li>Mood, genre, tempo, "Instrumental.", "Lyrics in English." go in 🎵 Song. They are put in front of scene 1's text for the LLM, so scene 1's text is only the scene. Empty = write the song in scene 1's text as before</li></ul>🎚 The MV's sound, and remaking only the sound (v7.0)<ul><li>🎚 Sound: 🔊 even loudness and its LUFS target, 🎛 EQ (low 120 Hz, mid 1.5 kHz, high 8 kHz, ±12 dB), 🏛 reverb (amount, length), ↔ stereo width, 🌅 fade in / out, ↺ back to as it is. The EQ and the reverb don't change the loudness (it is matched again) and the peaks stay under -1 dBFS. All at their defaults = the sound passes through untouched</li><li>It works on the finished video's soundtrack only, never on what H3 hears, and nothing is moved in time, so the picture stays in sync</li><li>🔁 Sound only: "🎚 sound only" in the 🔁 strip (or "Keep the picture, make only the sound again" in 🎚 Sound). Every scene of the kept run is read from its file (no H3, no decode, no YuE2) and the video is joined again with the new sound. Use it every time you only change the sound: with 🎲 random seeds an ordinary run draws a new picture too</li></ul>➕ More LoRAs and ⚡ one Turbo steps setting (v6.4 / v7.0)<ul><li>➕ More LoRAs (in 📦 Models): up to three LoRAs after the H3 model, for every scene, with ⚡ Turbo on and off. (none) or strength 0 = not used</li><li>⚡ Turbo steps (🔧 Details, default 4): one setting for the H3 steps with Turbo on, for every scene. It used to be a number inside each of the two H3 subgraphs, so the scenes could differ. With a 4–8 step Turbo LoRA, try 6–8 on one scene with 🔁 first. Turbo off still uses 20 steps, and the 💾 Civitai metadata gets the steps really used</li></ul>🎸🎻 The hands play along (v6.3)<ul><li>🎸 The fretting hand, timed: while the player strums or sings, "At 00:04.000 S1's fretting hand changes to the next chord shape" at each chord change of the score (a bassist: "moves to the next root note"), at most 6 a shot. In a solo, the hand moves along the neck where the melody jumps (guitar / bass: 5 semitones or more that don't come straight back; violin: when a note leaves the hand's position, first position G3 to B5), at most 4 a shot. Two shifts the same way less than 1.5 s apart are one move, said once</li><li>🎻 Bow and fingers from the note lengths (violin, viola, cello): quick notes (the fingers drop one after another, the bow short and quick, or smooth when slurred), note by note (the bow changes direction with each note, about every N s), held notes (one finger down with a slow vibrato, one slow bow). Slurs and ties are read from the score; an instrumental's melody is the Vocal voice alone</li><li>🎥 A close-up of the fingers when a solo really starts: "the camera moves in to an extreme close-up of S1's left-hand fingers on the fingerboard, the face out of frame", 2.5 to 4 s, then back out. One a shot, never where the shot or the next one is a close-up already, not in a scene's first 1.5 s</li><li>fret_cues / hand_closeup / bow_finger_cues (all on by default) are on the 🥁🎤 node and in ⚙ Settings</li><li>⚠ Strings set by hand that disagree with the audio: a 🎻 span dragged or deleted by mistake on the 🎞 Timeline (the violinist holds the violin down while it plays) is marked on the lane (orange / yellow) and in the 🎼 report, from 3 s a scene</li></ul>Note: H3 decides the fingers; correct fingering is beyond any video model today. The aim is a hand that moves when the music changes. Hands look best with ⚡ Turbo off (20 steps).🔧 Fixed<ul><li>🎻 An instrumental's melody (for the hand cues) is the Vocal voice alone; the accompaniment is no longer mixed in</li><li>🎻 With cue_detail "changes", a song whose violin rests at the very start stopped with an error ('rest')</li><li>🎸 A hand close-up was added at the start of a scene where the melody only went on from the scene before</li><li>🎞 The 🔁 strip shows its ✕ even when no run is kept yet, and ✕ also clears "🎚 sound only" (a leftover would stop the next ordinary run)</li></ul>Custom nodes in this download: ComfyUI-PromptSplitter v3.3, ComfyUI-ABCTiming v2.52, ComfyUI-ABCScoreIO v1.49 (all three updated) and ComfyUI-ABCInstrumentalize v1.5 (unchanged). Updating from v6.2? Replace the ComfyUI-PromptSplitter, ComfyUI-ABCTiming and ComfyUI-ABCScoreIO folders in <code>ComfyUI/custom_nodes/</code>, restart ComfyUI and reload the page (Ctrl+F5). Use the v7.0 workflow: it has the new 🎭 Cast &amp; song group, ➕ More LoRAs, 🎚 MV sound and ⚡ Turbo steps nodes. The v6.x workflows still open with these nodes.

onePromptOneSongATwoScene_v70.zip

v7.0

v6.2

v6.1

v6.0

v5.9

v5.8.1

v5.7

v5.6

v5.4

v4.1

v3.9

v4.0

v3.7

v3.6

v2.19

v2.09

Files with SHA256: 7c8f00115ea21fa158a7b2c3c63b4d3dadf1ee70d626957113b07d768933a832