Questions between the lines of the AI Selfie party..
ShippedWith.AI is a weekly roundup of interesting things built atop AI, paired with ideas for entrepreneurs and investors in the space. The AI market moves fast. Our goal is to help you understand and find inspiration in it.
Thought of the week
Actually doing a thing gives you a different, instinctual understanding than merely reading about it. In that vein, several things become clear after training DreamBooth on my face last week:
The Total Addressable Market of “better looking photos of yourself” is everyone.
The consumer interface will be Snapchat lenses, not prompts.
Prompts will be in the “advanced tab,” and you can sell yours for 99c. The ringtone market, but now we’re all ringtone makers.
Corporate headshot photography is finished.
To summon Don Draper: these things are time machines. There’s something about seeing yourself storm the beaches of Normandy that’s more than mere amusement. I’m not sure what it is yet.
This Week’s Links
If you’re on this list, I figure you’re already saturated with AI Selfie tweets. So I tried to add questions and comments below that I haven’t seen yet on Twitter. We’ll start as close as I can locate to where the frenzy began: on Reddit.
Fast-DreamBooth.ipynb by TheLastBen - AI selfies for free in Jupyter Notebook
(Reddit thread). Here’s the million dollar question: what stopped this notebook from being a Streamlit-style web app? The person that figures that out stands to drive a lot of business. Most people don’t know what a Jupyter notebook is. Of those that do, most would rather pay $40 than train a model themselves. I’ll try to chat with Ben and tell you what I find.
AvatarAIMe by Peter Levels - First to market a Stable Diffusion selfie generator.
Peter’s work is always a case study in implementing the essence and skipping the rest. In this case, the app is very nearly a Google Form. He may have even been generating photos semi-manually at the beginning. I think the takeaway is that some of these models are so good, the model really does sells itself — the profit margin comes from (1) making it easy to run, and (2) producing good results predictably. Platform builders, take note.
ProfilePicture.AI by Danny Postma - Stable Diffusion selfie generator, Round 2!
Second out of the gate by 24 hours. Expands to pet photos. Adds web polish that feels more SaaS. And that design difference provokes a question: do we know where the line between “fad opportunity” and “long-running business” is yet? A Google Form-style interface feels optimized for the former. A slick website feels optimized for the latter.
ImagineAI.me by Arib - AI Selfies for 1/3 the cost.. how does he do it?
Here’s the question that really has me scratching my head — how quickly will the cost basis of this change? Peter and Danny have grumbled about profit margin on Twitter, but here’s Arib wrapping Stable Diffusion for 1/3 the price ($10). A bit of EC2 dart throwing tells me you can get an instance to fine-tune Stable Diffusion for about $3/hour. Extrapolating from my own experience of about 2.5 hours, ImagineAI has got to be about the lowest price possible without infrastructure magic. Curious if the folks at Modal are going to use this as an opportunity to jump in with their new GPU offering.
PhotoAI.me by Seb Lhomme - AI Selfie packs, tuned for different occasions
Seb is experimenting with the idea of selling distinct “packs,” and if there’s a way to pad the fine-tuned AI margins, this is it. Tuning the model is the big fixed cost. Generating each pack is near zero marginal cost — probably within the minimum billable unit for your spot inference machine. The big question is whether prompts, or whatever replaces them, will remain proprietary enough to make this model sustainable.
That’s it!