seedance 2.5 is getting f*cking crazy for ai ugc
made an ai version of @brezscales getting stopped in Miami to explain the fiji water biz
the setup behind this:
> a photo for his appearance
> separate voice references for him and the interviewer
> a short video reference for the interview movements
> a car reference for the background
> a script broken into timed exchanges
giving each reference a specific job makes the prompt much easier to control.
the movement reference should guide gestures and microphone handling.
otherwise you can end up inheriting its camera moves and having everyone walk through the entire conversation.
same with audio. assign each voice to a named speaker, repeat that assignment beside their dialogue, and specify outdoor recording conditions.
for AI UGC, you can build a product angle into that conversation:
“what do you use for that?” → show the problem → explain the product → answer the obvious objection.
another example in case yesterday wasn’t convincing enough.
one real creator. one ai.
would you be able to tell which is which while scrolling?
this is why i wouldn’t write off seedance 2.0 yet.
one of these girls is real. two are ai.
pick the real one before you replay it.
everyone moved on to seedance 2.5 like 2.0 suddenly stopped working.
it didn’t.
and for short ai ugc variations, the cheaper model still deserves a place in your stack.
here’s how i’d pair it with gpt astra:
→ astra breaks down the reference: framing, gestures, timing
→ prepares the prompts and assigns each reference a job
→ operates the generator through your browser
→ seedance 2.0 creates the video
keep the structure. change one thing at a time.
new character.
new setting.
new opening line.
now you’re testing what actually changes the response instead of making five completely different ads and guessing why one worked.
2.5 can have the hype.
2.0 can still get the fucking work done.
drop “clone��, follow + repost and i’ll dm you the workflow.
one of these girls is real. two are ai.
pick the real one before you replay it.
everyone moved on to seedance 2.5 like 2.0 suddenly stopped working.
it didn’t.
and for short ai ugc variations, the cheaper model still deserves a place in your stack.
here’s how i’d pair it with gpt astra:
→ astra breaks down the reference: framing, gestures, timing
→ prepares the prompts and assigns each reference a job
→ operates the generator through your browser
→ seedance 2.0 creates the video
keep the structure. change one thing at a time.
new character.
new setting.
new opening line.
now you’re testing what actually changes the response instead of making five completely different ads and guessing why one worked.
2.5 can have the hype.
2.0 can still get the fucking work done.
drop “clone”, follow + repost and i’ll dm you the workflow.
this is f*cking insane.
gpt-6 astra ran this whole ad production through my browser
it found the references and pitched me the idea:
→ pov holding trenchies while walking toward an svj
→ cut inside the parked car
→ a short product pitch that fits an online workday
it analysed the motion clip, wrote the script, assigned the references and operated atlas cloud
seedance 2.5 generated the video you’re watching
the sauce is giving every reference one job:
product image → packaging
character image → appearance
motion clip → movement and timing
voice clip → delivery
then build the ad around one simple reason someone would use the product
show it. put it in context. make the point
that’s a repeatable structure for AI UGC
i put the setup, settings and a copyable starter prompt into a short guide
comment “gummies”, follow + repost and i’ll send it over
seedance 2.5 is getting f*cking crazy for ai ugc
made an ai version of @brezscales getting stopped in Miami to explain the fiji water biz
the setup behind this:
> a photo for his appearance
> separate voice references for him and the interviewer
> a short video reference for the interview movements
> a car reference for the background
> a script broken into timed exchanges
giving each reference a specific job makes the prompt much easier to control.
the movement reference should guide gestures and microphone handling.
otherwise you can end up inheriting its camera moves and having everyone walk through the entire conversation.
same with audio. assign each voice to a named speaker, repeat that assignment beside their dialogue, and specify outdoor recording conditions.
for AI UGC, you can build a product angle into that conversation:
“what do you use for that?” → show the problem → explain the product → answer the obvious objection.
Codex + GPT Image 2 for restaurant menus is such a good business idea
look at these. i'd want to order half the menu
you could build a system that:
→ finds restaurants in a city you choose
→ opens their websites and reads the menus
→ creates a visual menu preview
→ drafts an email with the mockup + a booking link
check the results, then send the previews
offer the full redesign, then ongoing updates for new dishes and seasonal menus
i put the build prompt, image prompts and outreach template into a guide
follow + repost + comment “MENU” and i'll send it over
Astra + GPT Image 2 is a ridiculous combo for restaurant visuals
the cheese pulling through the text is so good
you could automate a lot of this:
→ find local restaurants
→ pull their menus and dish photos
→ generate a few ad concepts
→ draft an email showing the results
review everything, then reach out
if they like it, offer a monthly batch for specials and menu launches
one example prompt in the comments ↓