Sjonker

ai filmmaker. I test what actually works in seedance, kling & co and share the exact prompts

[Sjonker @Sjonkerr]

same characters, same location, same lines.

first: i only typed the dialogue
second: i directed it like a real actor

night and day. what i put in the prompt 👇

[Sjonker @Sjonkerr]

@iamneubert 30 seconds, all those idioms and the same girl in one generation? Then the prompt is the real flex. Looking forward to the breakdown, I'm going to try it.

[Sjonker @Sjonkerr]

@CharaspowerAI Strong one. One thing I'd move: BREACH only shows up past the halfway mark of your prompt. Models follow the first ~20% most strictly, so I'd name the word and its material in line one. Dan Kieft also adds "one continuous virtual camera journey with no hard cuts" for type.

[Sjonker @Sjonkerr]

@umesh_ai can confirm, i measured it. a 6 sec draft cost me 18 credits, finalizing that exact take in 1080p was 96. so 4 drafts + 1 final ends up at less than half of 4 tries straight in 1080p. and you keep the take you liked

[Sjonker @Sjonkerr]

@CallMeAlgy @renoiseai @renoiseaijp @kfc cutting the best bits from a few takes is the right move, that's how most good ai stuff gets made. for keeping her identical across videos i use one sheet: big face close-up + the body without the head, on grey. way less drift than a full body shot with a tiny face

[Sjonker @Sjonkerr]

@NEXUS_TO_NOVA same, i block everything in blender first. what fixes headcounts for me: a top-down layout with a coloured dot per character, labelled as blocking reference for starting positions only. and if all six wear white with white hair it'll merge them, give each one a distinct detail

[Sjonker @Sjonkerr]

@second_bassoon tested this today with a villain scene. same lines twice, once just the dialogue, once with what he wants + a tone per line + beats between the lines. completely different scene. the directing does more than the words

[Sjonker @Sjonkerr]

setup if you want to try it:

both characters made with soul cast, the warehouse with soul location. then seedance 2.5, 720p, same 3 refs and the same camera line in both prompts

only thing i changed was the directing

[Sjonker @Sjonkerr]

same characters, same location, same lines.

first: i only typed the dialogue
second: i directed it like a real actor

night and day. what i put in the prompt 👇

[Sjonker @Sjonkerr]

what made the difference:

1. tell it what each character wants (he wants him to break)
2. a tone per line. quiet, then explode, then whisper
3. physical beats between lines. slam, flinch, the bulb swinging
4. give it 2 seconds of silence

the words were exactly the same

[Sjonker @Sjonkerr]

@chrisfirst 1.7% to almost half in under two years is crazy. soon "made with AI" will mean about as much as "made with photoshop"

[Sjonker @Sjonkerr]

@MadMax_Series Right? Most AI dialogue still sounds like a voice-over laid on top. This one sounds like the characters actually mean it.

[Sjonker @Sjonkerr]

The prompt structure:

"Shaky phone footage filmed by a passerby" first. The medium sells the realism.

Then 3 beats with seconds: roll in (0-2s), kickflip (2-4s), teens react (4-6s).

Sound written out: wheels on cobblestones, the pop of the trick, "NO WAY". No music.

[Sjonker @Sjonkerr]

Amsterdam grandmas don't care how long you've been skating.

AI video. Left: the cheap 480p draft. Right: the exact same take in 1080p.

What should she try next? 👇

[Sjonker @Sjonkerr]

Most people burn their budget finding the shot in full res.

Seedance 2.5 has a Draft mode:
1. Generate in 480p, fix camera, action and timing there
2. Finalize only the winner. It renders that same take in 1080p

This draft cost less than 1/5 of the final.

[Sjonker @Sjonkerr]

@zakburgers Before switching tools, check how you generate. Most of the spend is retries in full res. Seedance 2.5 has a 480p draft mode: find the shot there, finalize only the winner. In my test the draft cost <1/5 of the final.

[Sjonker @Sjonkerr]

@SelendisAI Learning taste from the cuts you make is the interesting part. Does it also learn from the cuts you undo? That's where a lot of editing decisions actually live.

[Sjonker @Sjonkerr]

@ghosstty_ The POV ones work because they read as filmed, not generated. Spelling out the medium does most of the work: "GoPro on a helmet, fisheye, wind noise, handheld shake" beats any amount of "cinematic, 8K".

[Sjonker @Sjonkerr]

@second_bassoon 5-7 tries per clip is exactly where Seedance 2.5's draft mode pays off. Do the tries in 480p, then finalize only the keeper in 1080p, it renders that same take.

In my test the draft cost less than 1/5 of the final.