Merci Grace

In 2021 I promised @l2k I'd work on whatever kept him up at night at @wandb. A few months later he said that thing was "sales." I'd never carried a quota in my life. Here's what happened next 🧵

@nabeel @kevintwohy Here's what I eventually got that I'm quite happy with. I did broad exploration with Midjourney and then shaped it with ChatGPT into both Default and Dark Mode. It's not pixel perfect but w/e. It'll be the first of my vibe apps to make it to the App Store, it's in review now.

yeah, I often find myself bouncing between 2-3 models and 3-4 interfaces. -Midjourney for style (and low adherence) if style is needed. -Ideogram for iconic work. -MJ for small area edit control, unless those are by discrete object then you want Reve. -Nano Banana for general instruction following, but only in language its trained on which you kinda have to know. -ChatGPT when very tight scene study is important like what kevin mentioned, but the style is horrid. So sometimes start there and then use an control model style transfered into a different scene to get the look. It's a lot. And no way an average user is going to pay for all that and try and figure all that out so it's frustrating. [proverbial it's early statement goes here]

I've spent hours and probably $50 trying to coax three different image generators to make a simple but specific app icon. I'm about to hire a human being, this this ridiculous. Here's an example from Midjourney. 🙃

@kevintwohy @nabeel Yeah, it's totally magical if you have a general wish for an image; it's still clumsy if you know what you want at a high degree of granularity/taste. It'll get there, no doubt.

@nabeel @merci In particular most of the image models are *very* bad at verbal instruction around “camera” (angle, composition, zoom, framing). Needs a new interface.

@awwstn @nabeel @garrytan This chart and its underlying data raise more questions than are answered by the editorial here, though that ChatGPT is stickier than the others is probably correct. Whether these are faithful J curves and for which defs of users across 4 diff apps, less than clear.

@nabeel @garrytan isn’t this stronger than a smile because it never dips?

I don’t think a tech product has ever had a retention curve that inflects upwards. We are so early

@nabeel Also, good luck getting it to remove the rounded corners and white border. This is why you can't tell any of the models that it's an app icon; the platform requires an edge bleed 1024x1024 image and if you have a border/rounded corners already it looks bad.

@nabeel A lot better! It sort of conforms to the form, but it’s not actually ✨good✨ enough to use. In my experience once I get something that adheres to the form, any further requests about style then obliterate the form.

@faraz_r_khan Wild how hard it is to get a specific perspective and item set.

@merci Have you tried ChatGPT or nanobanana? Here’s an image using shapr3ds integrated cad to image which is actually true to the cad model. It’s definitely hard to make it do exactly what you want though.

@meetbryce These are not the sum total of attempted prompts.

@merci consider saying it’s a logo. how would it know otherwise?

ChatGPT had the best output but cannot be coaxed into getting the perspective correct either.

Here's the latest from Google... 🍌

It's so fun to build things right now, I really needed this shot of absolute joy from the outside world.

@matt_slotnick Thank you! 😎

@merci welcome back!! (to the future)

Working on an app from a Philz like it's 2015. Except that it's just me, Claude and Ghostty.