04_
SÉJOUR EN ITALIE
01_ THE IDENT
02_ HOW IT WAS MADE
I started seeing people on social media remake scenes from TV shows with the new video models, and I wanted a way in. MiniMax-H3 had just come out, so I built a small UI around its API: text, frames, and references go in, takes come back, and the ones worth keeping stay side by side. The first thing I wanted to remake was the HBO intro. I love HBO shows, The Sopranos is my favorite show of all time, and the static that opens every episode is a classic.
I make videos for fun, and I wanted this one as the intro to a video I cut from a trip I took to Italy. So the ident goes the way it always has, the black, the line of light, the static, the hiss turning into the chord, and clears into Séjour en Italie where the logo would be. The original went in as a video reference with its sound, plus one keyframe with the title in place of the logo. It took a run of takes to get the order right, and the one above is the last of them, rendered at 2K and cut to the original soundtrack.
The inputs
Three things went into the model and they never changed: the original ident as the video to edit, with its own soundtrack copied so the hiss and the chord land where they always have; and one keyframe, the ident’s frame at 3.8 seconds with the HBO mark replaced by the title, so the model had a picture to arrive at. The prompt said the rest: black, a line of light, static, the title surfacing out of it, the static draining away, the title alone, the fade.
This is the shot as it sits in the studio. The keyframe is an edited keyframe, the ident is the edit source, the audio is a copy. Everything that follows was a change to one of those three or to the words around them.
Left, the shot as sent. Right, Video 1: the ident itself, six and a half seconds, sound on.
Video 1 · Edit source
the line of light at 2.1 s
Video 1 · Edit source
the frame at 3.8 s, the logo in the static
Picture 1 · Edited keyframe
the same frame, the title in place of the logo
Picture 2 · Edited keyframe, from take 10 on
the title alone, white on black, at 5.0 s
The first nine takes
Most of the first nine takes found the title. What they could not agree on was the way there. The first had the timing wrong: it cleared to the title by three seconds and then fell back into static. Others never left the static, or skipped it, or brought the title in late. The eighth got there, but the model invented a sound over the black opening, where the ident is silent, and it had to go. The ninth, rendered at 2K with sound, had the whole arc: black, the line, the static, the title clearing out of it, the title alone.
It had one thing wrong. Between 2.8 and 3.2 seconds the HBO mark ghosts through the static before the title surfaces. The model had reproduced what its reference showed, logo included, and the keyframe only told it what 3.8 seconds should look like, not what 3.0 should not.
Take 1 · Rejected
The timing is off. Clears to the title by three seconds, then falls back into static.
Take 8 · Rejected
A sound the model made up plays over the silent black opening. Goes dark early too.
Take 9 · Rejected
The whole arc, at 2K with sound. The HBO mark ghosts through at three seconds.
The ghost, and what it took
Then I went back for it. The first idea was the obvious one: take nine as the source, a prompt that defines the ghost as a subject and asks for its removal, everything else preserved. Take 10 came back with the ghost exactly where it was. The model keeps what its edit source shows; it will not erase something it can see. So the next take made sure it could not see it. Take 11 got only the first 2.6 seconds of the ident, before any letters form, as a motion reference, with the keyframe and the copied soundtrack. No ghost, and no letters of any kind, but with the source’s clock gone the model re-timed the whole ident: the line came early, the drain came early, the title emerged soft.
Takes 12 and 13 put the clock back without putting the logo back. The edit source became the ident itself with every frame from the logo’s first appearance replaced by the ident’s own clean static, then the drain, then black, and a second keyframe pinned the white title at five seconds. The look and the emergence were right. What the model still would not do was hold the two-second black opening; every full generation compressed it to under a second. Take 14 is 13’s exact request at 2K, and the fix for the clock was not another render. I cut inside its holds, the black, the static, the white, and laid the original soundtrack under it, so every event lands on the ident’s own timing.
Three seconds in, at twice size. Left, take 9: the HBO mark showing through before the title. Right, take 14: the title surfacing, nothing else.
Take 10 · Rejected
Take 9 as source, “erase the ghost”. The ghost stayed. The model keeps what its source shows.
Take 11 · Rejected
Only the first 2.6 s as a motion reference. No ghost, but the whole ident re-timed itself.
Take 13 · Rejected
The ident with its logo frames replaced by clean static. Right look, black opening compressed.
Take 14 · In the film
Take 13’s request at 2K, cut inside its holds to the original clock, the soundtrack laid under.
The prompt, as sent · take 10, the erase that did not work
subject_definitions: <Video 1> is the source video for the target video edit. It is a 16:9 analog television ident, 6.6 seconds long: black; a thin horizontal line of white light across the middle of the frame; the line opens into full-frame analog television static; the words "SÉJOUR EN ITALIE" surface out of the static in bold, wide, sans-serif capitals on two centered lines; the static drains away to black; the title stands alone in white on black; the title fades out. <Video 1> carries its own soundtrack, a static hiss that turns into a sustained chord. <Subject 1> is the analog static of <Video 1>: fine grey-white video noise filling the whole frame, with no letters, shapes, or marks in it. <Subject 2> is the title of <Video 1>: "SÉJOUR EN ITALIE", two lines of bold wide sans-serif capitals, centered, first raised faintly inside the static and finally solid white on black. <Subject 3> is a mistake in <Video 1>: between about 00:02.700 and 00:03.300, a large ghost of three block letterforms with a ring inside the third one shows through the static, above the position of the title, before "SÉJOUR EN ITALIE" has surfaced. summary: The target video is an edited version of <Video 1>. [video editing] The single change is the removal of <Subject 3>: the ghost letterforms are erased, so from 00:02.400 until the title begins to surface the frame shows nothing but <Subject 1>, plain static with no letters, ring, emblem, or watermark of any kind, and the only text that ever appears anywhere in the video is "SÉJOUR EN ITALIE". Everything else in <Video 1> is preserved exactly: the black opening, the line of light, the static texture, the timing and manner of the title surfacing, the drain to black, the white title, the fade, the framing, the camera, and the soundtrack. retention_analysis: <Video 1>: fully_preserved, its timeline, framing, effects, timing, and camera unchanged apart from the erased <Subject 3>. <Subject 1>: fully_preserved. <Subject 2>: fully_preserved. <Subject 3>: partially_preserved, erased and filled with <Subject 1>. <Video 1> audio: fully_copy. detailed_description: A vintage analog television ident, identical to <Video 1> frame for frame except that the ghost letterforms are gone. [Shot 1] Black. A thin horizontal line of white light appears across the middle of the frame and opens up and down into full-frame analog static (<Subject 1>). [Shot 2] At 00:02.400 the static fills the frame and holds. Nothing shows through it: no letters, no ring, no emblem, no watermark, only noise. [Shot 3] At 00:03.100 "SÉJOUR EN ITALIE" (<Subject 2>) begins to surface out of the static exactly as it does in <Video 1>, first as a faint raised impression in the noise and then clearer, centered, on two lines. [Shot 4] At 00:04.600 the static drains away to black and the title stands alone in solid white on black, as in <Video 1>. [Shot 5] At 00:06.000 the title fades out to black, as in <Video 1>. The requested removal is the sole changed element; the source framing, background, effects, timing, and camera remain preserved. overall_soundscape: The complete soundtrack of <Video 1>, copied unchanged: a rising static hiss that resolves into a sustained chord as the title appears, then decays with the fade. No added sound. non_diegetic_music: The sustained chord already in <Video 1>, copied unchanged. No other music.
The prompt, as sent · take 14
subject_definitions: <Video 1> is the source video for the target video edit: a 7-second analog television ident. Black for two seconds; at 00:02.100 a thin horizontal line of white light appears across the middle of the frame and opens up and down into full-frame analog static, which fills the frame by 00:02.500 and holds; at 00:04.000 the static drains away to black over 0.8 seconds; from 00:04.850 the frame is black to the end. The space where a title belongs is empty in <Video 1>: plain static, then plain black. <Subject 1> is the analog static of <Video 1>: fine grey-white video noise filling the frame, flickering, with nothing in it. <Picture 1> is an edited keyframe: <Subject 1> with the title "SÉJOUR EN ITALIE" raised inside it, bold wide sans-serif capitals on two centered lines, the letters slightly lighter than the noise with a soft glow along their edges. The target must match <Picture 1> exactly at 00:03.800. <Picture 2> is an edited keyframe: the same title "SÉJOUR EN ITALIE", solid white, standing alone on black, centered. The target must match <Picture 2> exactly at 00:05.000. <Subject 2> is the title "SÉJOUR EN ITALIE" as shown in <Picture 1> and <Picture 2>: the only text that ever appears. <Audio 1> is the complete 7-second soundtrack, to be copied unchanged: digital silence until 00:02.100, then a static hiss that rises as the line opens into static, then a sustained synthesizer chord that swells as the title surfaces and holds to the end. Its timing is the timing of the picture: the line of light appears at the instant the hiss begins. summary: The target video is an edited version of <Video 1>. [video editing + keyframe completion + audio reuse] The single change is that the title "SÉJOUR EN ITALIE" (<Subject 2>) is added where <Video 1> is empty: it surfaces out of the static between 00:02.750 and 00:03.250, is fully formed in the static by 00:03.800 as in <Picture 1>, stays through the drain, stands alone in white on black from 00:04.850 as in <Picture 2>, and fades out with the frame between 00:06.300 and 00:06.800. The timeline of <Video 1> is kept to the frame: nothing happens before 00:02.100, the static fills the frame at 00:02.500, the drain runs from 00:04.000 to 00:04.850. Every other element of <Video 1> is preserved exactly: the black opening, the line of light, the static and its texture, the drain, the black, the fixed camera. The sound is <Audio 1>, copied unchanged and locked to the picture. No other text, letters, emblems, or logos appear at any point. retention_analysis: <Video 1>: fully_preserved (timeline, framing, texture, timing, camera). <Subject 1>: fully_preserved. <Picture 1>: fully_preserved, matched exactly at 00:03.800. <Picture 2>: fully_preserved, matched exactly at 00:05.000. <Subject 2>: fully_preserved, the only added element. <Audio 1>: fully_copy. detailed_description: A vintage analog television ident, identical to <Video 1> in every respect except the added title, and cut to <Audio 1>. [Shot 1] Black and silent for the first 2.1 seconds, exactly as <Video 1> and <Audio 1>. [Shot 2] At 00:02.100 the hiss of <Audio 1> begins and, at the same instant, a thin horizontal line of white light appears across the middle of the frame and opens up and down into full-frame static (<Subject 1>), filling the frame by 00:02.500, exactly as <Video 1>. [Shot 3] At 00:02.750 "SÉJOUR EN ITALIE" (<Subject 2>) begins to surface out of the static as the chord of <Audio 1> swells, first as a faint raised impression in the noise, becoming clear and fully formed by 00:03.250, and holds in the static so that at 00:03.800 the frame matches <Picture 1> exactly. [Shot 4] At 00:04.000 the static drains away to black over 0.8 seconds, exactly as <Video 1>, while the title stays in place and turns solid white as the noise leaves it, so that at 00:05.000 the frame matches <Picture 2> exactly: the title alone, white on black. It holds while the chord sustains. [Shot 5] At 00:06.300 the title fades out to black over half a second; the frame is black to the end while the chord continues. Fixed camera, no movement, no color, no other elements. overall_soundscape: <Audio 1>, copied unchanged, is the entire soundtrack: silence over the black opening, the static hiss beginning exactly with the line of light at 00:02.100, the hiss giving way to the sustained chord as the title surfaces, the chord holding to the end. No added sound. non_diegetic_music: The sustained synthesizer chord contained in <Audio 1>, copied unchanged. No other music.
What carried over
Three rules came out of this and every shot since has run on them. Remove a thing from the model’s inputs rather than asking the prompt to suppress it. Keep the source’s timing in the inputs, or the model will make its own. And take the soundtrack from the source file, not from the model, then cut the picture to it. The studio changed with the work too: the reference roles, the audio copy, the keyframe pinned to a timecode and the desk that lines a sound up with a frame were all built for this ident before they were used on anything else.
