“I changed a word and tried again. I don’t actually know what was wrong.”
That is the loop.
Something fails, you adjust a word, you spend another generation, and you still do not know what happened. Do it four times and you have paid four times for one shot.
Your model is not broken. It is doing something specific, and that something has a name and a cause. This manual is 105 of them.
105findings
89measured, not guessed
37fixes that cost nothing
29providers
Version 1.28, updated 11 October 2026. When we learn something new, it goes in the manual.
Three of the 105, in full.
Not excerpts and not a teaser. These are complete entries, the same three the notebook turns to. If they are not worth $19 to you, the other 102 will not be either. Better you find that out now.
Seedance 2.5costs a re-rollmeasured
The label came back as melted letters. The clock on the dashboard is scribble. You re-roll and get different scribble.
What is actually happening
Type that sits small in the frame does not resolve. The same words render cleanly when the camera is close enough for them to be large.
The fix
Reframe so the words fill a good part of the frame - a close-up of the phone, not the phone across the room. If the words cannot be large, leave them out of the generation and add them in your editor over the finished clip, which costs no generation at all.
On another platform
This is how the model draws, so we expect it to travel. But it is what we saw on one platform, not a law. The platform sets the output size: if yours sends 1080p where ours sent 720p, borderline type may survive. Check what resolution your platform really asks for.
Measured, on fal. Measured 19 Sep 2026 across $16.10 of takes on fal: the phone close-up came back legible on every attempt; the dashboard clock and the distant phone screen came back as gibberish on every attempt.
Seedance 2.5costs a re-rollmeasured
You wrote "locked off, no camera movement" three times and the camera still floats around. But the change you asked for at 2 seconds does arrive at 2 seconds.
What is actually happening
A timecode in the prompt does place an event at that second. A request to hold the camera still is ignored, however many times you write it.
The fix
Write every beat as a different move - "0-2s slow pan left, 2-4s push in to medium" - never as a move against stillness. For a beat that truly has to be locked, make that one shot on seedance-1.x with its camera-lock switch (camera_fixed) turned on, and cut it in.
On another platform
The timecode behaviour and the refusal to hold still look like how the model works, so we expect them to travel. But this was one 4-second take on one platform, so read the 80% figure as one measurement, not a rate. The camera-lock switch only helps if your platform offers seedance-1.x with that switch showing. Many apps do not show it.
Measured, on fal. Measured 20 Sep 2026 on fal, $1.84, one 4s take scored frame by frame: motion jumped 0.493 to 0.890 between 2.00s and 2.17s at exactly the stated second, and the second half ran 2.63x the first. "LOCKED OFF", stated three times, was ignored: motion started at 0.17s, and frames at 0.00s and 1.90s differed by 32.8/255 where the prompt said identical.
Every providerfree fixdocumented
Same model, same clip, and one platform's bill is a third bigger than the other's.
What is actually happening
The same model is resold at different per-second rates depending on where you buy it.
The fix
Price the clip on every service you already use before you commit a batch. For us, Seedance 2.5 at 1080p was $0.68 a second on Runway and $1.1372 a second on fal - about $14 apart on a single 30-second take. Runway adds 15 credits a second when you attach a reference video, so price a job with many references again rather than assuming the saving holds.
On another platform
Resale prices move, and these were read on 27 Aug 2026. A subscription platform hides the per-second rate behind credits, so the comparison there is credits per clip, not dollars per second. We priced fal and Runway only; we did not price Higgsfield, OpenArt or the provider apps for this model beyond one Higgsfield data point.
Documented, on fal and Runway, read from each one's published developer docs. Both sides checked in our 27 Aug 2026 Runway audit against Runway's published developer docs, with Runway credits at $0.01 each: Seedance 2.5 720p $0.30/s on Runway vs $0.4730/s on fal, 1080p $0.68/s vs $1.1372/s. Same table: Veo 3.1 without audio $0.20/s on Runway vs $0.40/s on fal.
Why it says where every finding was measured.
We measured on fal and on Runway. You may be on Higgsfield, OpenArt or a provider’s own app, where access and limits differ: the same model can refuse a face on one platform and accept it on another. So every finding says where it was measured, and every finding carries an On another platform line telling you what could be different for you and how to check in about ten seconds. A manual that claimed to cover every platform would be one you could disprove before lunch.
Cost is counted in re-rolls, not dollars, for the same reason. An 8-second 720p Seedance clip is about $3.78 bought direct and about $2.55 on a Higgsfield plan where credits expire monthly. One dollar figure would be wrong for nearly everyone. A wasted generation is a wasted generation everywhere.
Buy it if
- You generate AI video and re-roll more than you would like
- You want to know why it failed, not just try again
- You pay per generation, or burn a monthly credit allowance
Do not buy it if
- You want prompts to copy: there are none in here
- You have never generated a video and want a beginner’s guide
- You need guaranteed behaviour on a platform nobody has tested
Questions.
Will these work on my platform?
Some will, some may not, and each entry tells you which. Every finding says where it was measured and what could differ for you. That is the honest answer, and it is more useful than a confident one that turns out to be wrong.
What does “measured” actually mean?
We ran it and have the number or the price. 89 of the 105 are measured. The other 16 are read from a provider’s published settings or its own documentation, and the entry says so instead of dressing one up as the other.
Is this a prompt pack?
No. There are no prompts to copy. It is 105 findings with their causes, and in 37 cases the fix does not involve generating anything.
What if it is not useful?
Email me within 14 days and I refund it in full: one email gets you the whole $19 back. If you are in the EU or UK you also have statutory rights this does not affect.
How do I get it?
The download is on the page you land on straight after payment, with a link to that page you can bookmark and come back to. Yours to keep, and it works offline.
This manual diagnoses. I also make the work.
I design and build websites for businesses, and I make AI video and motion graphics for brands and products. The same measuring that went into this goes into the work. Tell me what you are trying to make and what it is for, and I will tell you whether it is worth making and what it would cost.
japhet@promptlenshq.com · See the websites
