kelly99
Joined 1 year ago
Karma
1
Submissions
10
Member since
Oct 13
Recent submissions
Comment · 8 months ago
Comment · 11 months ago
I’m a solo indie dev who has been building small AI-video tools for the past year. This week, I shipped… probably my strangest project so far:
a Sora 3 site — even though Sora 3 doesn’t exist yet. Project: sora3ai.io
To be clear: I don’t have access to Sora 3. No one does. It hasn’t launched.
All I did was build the interface, the workflows, and the “shell” around what I think people will want once it comes out. It’s basically a speculative product — like building a case before the phone is released.
Why did I do this?
A few reasons, none of which I feel fully confident about:
1. The ecosystem is moving faster than the models themselves.
Users already search for “sora 3 generator”, so wrappers appear before the actual API.
2. Indie developers don’t get access early, so building in advance feels like the only way to stay alive.
Otherwise, by the time APIs open, the big players have already taken the top 10 Google spots.
3. I’m trying to figure out whether “being early” actually matters — or whether this is all pointless.
My Sora 2 site got traffic but weak conversion. My storyboard generator had interest but little retention. Being early didn’t lead to product-market fit.
4. Is this ethical? Is it stupid? Or is it just how the AI ecosystem works now?
I genuinely don’t know.
Some questions I’d love to hear thoughts on:
* Is it reasonable to build tools for models that aren’t released yet?
* Does this create user confusion, or is it harmless speculation?
* As a solo dev, is “being the first wrapper” actually a viable strategy?
* Or is this whole AI-tools race destined to be a zero-margin commodity market?
I’m not trying to promote the site — it barely does anything yet. I’m more interested in whether this kind of “pre-emptive product building” is smart, predatory, or just inevitable.
Curious how people here think about this trend. Happy to share data, mistakes, or the weird SEO patterns I’ve been seeing.
— A slightly confused solo founder trying to keep up with a model that hasn’t shipped
Comment · 11 months ago
I’ve been running a small side project that needs fast and reliable image generation.
After trying SDXL/SD3, Flux, and a few closed APIs, I tested the new NanoBanana Pro models—and it kind of changed my whole setup.
A few things stood out:
1. It follows prompts unusually well. Not “creative drift”, not hallucinating weird hands, just… does what you ask. This alone cut my retries by almost half.
2. It’s really fast. On 1024×1024 I consistently see ~2 seconds, sometimes less. I didn’t expect that. It actually changes how you design the UI.
3. The default style is clean and usable. Not too glossy, not too anime, not too uncanny. Just realistic enough for product photos and people shots.
4. Cheaper in practice. Because it needs fewer retries, the cost per accepted image was lower than SDXL in my tests (1.3 vs 2.2 gens per final result).
I ended up building a small wrapper around it to normalize prompts, clean params and handle retries. If anyone’s curious, you can search “nan0banana” — I’m not posting a link so it doesn’t look spammy.
Curious if anyone else here has played with NanoBanana Pro. How does it compare to what you’re using?
Comment · 1 year ago
Hey HN,
I’ve been spending the past month diving into AI video generation — not just using models, but trying to understand the actual constraints behind them. After prototyping a small Sora-style generator on my own, I started to notice a few deeper patterns about the industry that I wanted to share and get feedback on.
1. AI video tools aren’t limited by “models”
Most of the friction today isn’t about model quality:
region-locked access
invite-only rollouts
heavy watermarking
friction in basic usage
short duration limits
no multi-scene support
pricing opaque or unsuitable for small creators
The technology is improving fast — but the accessibility layer hasn’t caught up.
This is why the majority of creators (especially small merchants, indie filmmakers, TikTok sellers, UGC creators) still can’t practically adopt AI video at scale.
2. Multi-scene generation is the “real moat”
Most models can do a single beautiful 2-4 second shot.
But real use cases — ads, storytelling, product demos — need:
shot transitions
visual consistency
character identity retention
stable camera paths
narrative structure
The real challenge is not “make a clip”， but “make a sequence”.
That’s where pipelines, not models, matter.
3. The real bottleneck is temporal coherence
From my experiments, the hardest problems aren’t fancy effects — they’re the boring ones:
slight drift in character identity
physics mismatch between shots
exposure shifts
motion jitter at boundaries
model choosing different “interpretations” each time
There’s no perfect solution yet. Some combination of:
prompt redistribution
style anchors
conditioning
intermediate frames
shot graphs
works “okay”，but there’s huge open research space.
4. Small creators care less about model elegance — more about “does it work for my product?”
This surprised me.
I talked to some merchants and small creators. What they wanted wasn’t:
“best model”
“highest fidelity”
“latest architecture”
They asked for:
no watermark
9:16 format
product-handheld shots
consistent 20–25s video
don’t make me wait
just give me something I can post today
It’s a very different set of priorities than what model researchers focus on.
5. The infra is the unsung hero
Most public discussions focus on models, but from building my prototype I realized:
async queues
model switching
fallback logic
caching policies
GPU scheduling
latency constraints
matter far more for practical AI video creation than architecture diagrams.
Without good infra, even the best models feel unusable.
A prototype I built while exploring these ideas
As a way to understand these bottlenecks more concretely, I built a small prototype called Saro2.ai — basically an experiment in:
10s cinematic clip generation
25s multi-scene “storyboard” generation
attempts at shot consistency
simple scene → shot graph
a multi-model backend with light scheduling
It requires login (to control compute use), but I’m mainly sharing it as an example of the things I’m testing, not trying to “launch a product”.
Here’s the link if anyone wants to see how it behaves: https://saro2.ai/
What I’m hoping to learn
If you’ve worked on:
temporal modeling
multi-scene pipelines
conditioning
generative video infra
shot consistency strategies
I’d love to hear your perspective.
Especially curious about:
what people think the real frontier is
what “must solve” engineering problems exist before AI video is truly usable
whether multi-scene consistency is solvable with heuristics or requires new architectures
Happy to share more details about the pipeline or what didn’t work.
Thanks for reading — and I’d appreciate any thoughts from people working in (or following) this space.
For a while, most outputs felt like interesting demos. Recently, models like Seedance have become much stronger at motion, consistency, and prompt adherence, to the point where AI video is starting to feel usable for real production workflows (ads, social clips, product content), not just experimentation.
That shift is why I built Veemo.ai: a workflow layer to actually use these models efficiently.
Veemo combines text-to-video, image-to-video, and video-to-video in one place, supports multiple models (including strong newer ones like Seedance), and is focused on fast iteration rather than one-off generations.
I’m especially interested in feedback from people using AI video in real work: - What still makes these tools hard to trust? - Where do you lose the most time in the workflow? - What would make you pay for a tool like this?
Happy to answer questions about reliability, model tradeoffs, and pricing.