the author mentions they’re using haiku for the model, but I wonder if the travel query -> place transform could be done with a tiny local model.
But that’s the bitter lesson I guess, unless there’s a reason to go fully local (extreme privacy concerns, offline use, etc) training a custom model just isn’t worth it over using some cloud api offering, even if it is orders of magnitude more compute that _somebody_ is paying for in the end
Couldn't you grab the image from the Wikipedia page of that location? For example searching for "Deadvlei" from the blog post gives this which has a photo:
The annoying thing with Unsplash is that it sounds like a really permissive license, but the moment you try and do it at scale you're no longer in compliance with their terms. Also their definition of not building a competing service is really broad. Maybe i'm being overly cautious but I get eeked out by all that.
The headline (currently: “Trying to craft AI images that are worth displaying to end users”) is misleading and changed from the original. Author isn’t crafting any AI images; they’re using AI in tandem with manual work to help choose from a set of human-authored images.
Ok! that was my attempt to avoid linkbait and make the title less provocative (submitted title was "How to design an AI app with a sense of taste"). But I missed the mark this time, so have reverted the title to the article's own headline, except I'm not going to keep the word 'beautiful' up there since that would be certain to provoke shallow objections.
The title on HN is incorrect/misleading, they are not generating AI images. They are hand curating a database of images by location and using an LLM to pick the pictures.
Speaking of "crafting", I think this is the perfect word to describe something more than "prompting".
It's extremely hard to block out a scene with just words, eg. "rotate hand 45 degrees, stand perpendicular to the column, shadows from light source 60 degrees above horizon, large box in front of chest, approximately 2 feet wide", etc.
Image-to-image, ControlNets, previz-to-final, etc. are the way to go, and I'm convinced this is the core interface for image and video creation. Text prompts will get you a coarse grained first approximation, which you then visually adjust to your exact needs with UI/UX-first models.
I built an intentional "crafting" engine so people could mold images like clay, with full intention:
This is really early days though. I expect more tools and models to enable you to fully manipulate everything first-class, in 2d/3d. As if everything in an image were mutable.
As a film director, this is really exciting stuff.
I realize dang has already tagged this (and I get it) but I don’t know who is actually benefiting from this. This is deceptive as it doesn’t represent a traveler’s actual/expected experience and just adds more slop to the pile. This is a case where you should pay actual photographers to take verified pictures of the places you are recommending people visit. Why would someone cheapen their brand like this?
If you read the article you'll see it's not about generating AI slop images:
> Take a freeform query (like ‘sfo->jfk’) and turn it into a ‘place’
> Build a database of ‘places’ -> pictures
> Build a software system that can take a ‘place’, look it up in a database and spit out the right picture – even if that ‘place’ isn’t in the database
But that’s the bitter lesson I guess, unless there’s a reason to go fully local (extreme privacy concerns, offline use, etc) training a custom model just isn’t worth it over using some cloud api offering, even if it is orders of magnitude more compute that _somebody_ is paying for in the end
https://en.wikipedia.org/wiki/Deadvlei
[sorry I messed up with that title! perils of not reading the articles closely]
It's extremely hard to block out a scene with just words, eg. "rotate hand 45 degrees, stand perpendicular to the column, shadows from light source 60 degrees above horizon, large box in front of chest, approximately 2 feet wide", etc.
Image-to-image, ControlNets, previz-to-final, etc. are the way to go, and I'm convinced this is the core interface for image and video creation. Text prompts will get you a coarse grained first approximation, which you then visually adjust to your exact needs with UI/UX-first models.
I built an intentional "crafting" engine so people could mold images like clay, with full intention:
https://github.com/storytold/artcraft
This is really early days though. I expect more tools and models to enable you to fully manipulate everything first-class, in 2d/3d. As if everything in an image were mutable.
As a film director, this is really exciting stuff.
If you show something photorealistic and AI generated, what is shown is simply an illusion.
If you use a cartoon style maybe they can work because the user will immediately understand what is shown is not a photograph.
https://news.ycombinator.com/newsguidelines.html
> Take a freeform query (like ‘sfo->jfk’) and turn it into a ‘place’
> Build a database of ‘places’ -> pictures
> Build a software system that can take a ‘place’, look it up in a database and spit out the right picture – even if that ‘place’ isn’t in the database