Comment by neom
14 hours ago
Here is my regular "hard prompt" I use for testing image gen models:
"A macro close-up photograph of an old watchmaker's hands carefully replacing a tiny gear inside a vintage pocket watch. The watch mechanism is partially submerged in a shallow dish of clear water, causing visible refraction and light caustics across the brass gears. A single drop of water is falling from a pair of steel tweezers, captured mid-splash on the water's surface. Reflect the watchmaker's face, slightly distorted, in the curved glass of the watch face. Sharp focus throughout, natural window lighting from the left, shot on 100mm macro lens."
google drive with the 2 images: https://drive.google.com/drive/folders/1-QAftXiGMnnkLJ2Je-ZH...
Ran a bunch both on the .com and via the api, none of them are nearly as good as Nano Banana.
(My file share host used to be so good and now it's SO BAD, I've re-hosted with them for now I'll update to google drive link shortly)
I mean, your prompt is basically this skit: https://www.youtube.com/watch?v=BKorP55Aqvg ("The Expert" 7 red lines: all strictly perpendicular, some with green ink some with transparent ink)
I couldn't imagine the image you were describing. I've listed some of the red lines with green ink I've noticed in your prompt:
Macro Close Up - Sharp throughout
Focus on tiny gear - But also on tweezers, old watchmakers hand, water drop?
Work on the mechanism of the watch (on the back of the watch) - but show the curved glass of the watch face which is on the front
This is the biggest. Even if the mechanism is accessible from the front, you'd have to remove the glass to get to it. It just doesn't make sense and that reflects in the images you get generated. There's all the elements, but they will never make sense because the prompt doesn't make sense.
The last point (reflection by front glass versus mechanism access so no front glass) is the only issue I see with it. Other than that I can easily visualize an image that satisfies the prompt. I think that the general idea is a good one because it's satisfable while having multiple competing requirements that impose geometric constraints on the scene without providing an immediate solution to said constraints as well as requiring multiple independent features (caustics, reflections, fluid dynamics, refraction, directional lighting) that are quite complicated to get right.
To illustrate that there aren't any contradictions (other than the final bit about the reflection in the glass). Consider a macro shot showing partial hands, partial tweezers, and pocket watch internals. That's much is certainly doable. Now imagine the partial left hand holding a half submerged pocket watch, fingertips of right hand holding front half of tweezers that are clasping a tiny gear, positioned above the work piece with the drop of water falling directly below. Capture the watchmaker's perspective. I could sketch that so an image model capable of 3D reasoning should have no trouble.
It's precisely the sort of scene you'd use to test a raytracer. One thing I can immediately think to add is nested dielectrics. Perhaps small transparent glass beads sitting at the bottom of the dish of water with the edge of the pocket watch resting on them, make the dish transparent glass, and place the camera level with the top of the dish facing forward?
https://blog.yiningkarlli.com/2019/05/nested-dielectrics.htm...
A second thing I can think to add is a flame. Perhaps place a tealight candle on the far side of the dish, the flame visible through (and distorted by) the water and glass beads?
Without the last point with the watch glass it is also easier to imagine for me. Still, you'd have to be selective.
Do you want it to actually look like macro photography (neither of the generated images do)? Then you can't have it sharp throughout and you won't be able to show the (sharp) watchmakers face in a reflection because it would be on a different focal plane.
Dropping the macro requirement, you can show a lot more. You can show that the watchmaker is actually old, you can show the reflection, etc.
Something has to give in the prompt, on multiple of the requirements. The generated images are dropping the macro requirement and are inventing some interesting hinging watch glass contraptions to make sense of it.
1 reply →
Yeah I dunno bud, I have a degree in film and three Emmy awards for technical production (an expert), I could shoot that prompt (unlike the so called "expert" in the skit). Canon EF 100mm Macro USM at f32 should be able to produce that, focus doesn't need to imply aperture, and a quick google search shows me there are loads of front gear pocket watches available. Also it produced something very clearly not shot with a 100mm anyway, as the telephoto compression is wrong.
Yeah I dunno bud, I've watched a few watch repair videos on youtube and have seen macro photography which other people did.
Sure there are pocket watches where the movement is visible from the front (you'd still likely service them from the back, but alas). Even if you'd do service from the front where the glass is, you'd still have to remove it to drop in a gear.
Anyway, I think that we aren't really talking about the same thing. I'm nitpicking your prompt while you constructed it to mostly see the performance of the model in novel situations and difficult lighting and refraction environments. And that's fair.
How satisfied are you with the generated image results? What would you do different when shooting this proposed scene yourself?
1 reply →
Why would you consider this a good prompt?
My observations have been that image generation is especially challenged when asked to do things that are unusual. The fewer instances of something happening it has to train on, the worse it tends to be. Watch repair done in water fits that well - is there a single image on the internet of someone repairing a watch that is partially submerged in water? It also tends to be bad at reflections and consistency of two objects that should be the same.
Looks like your image host has rate limited viewing the shared images, wanted to give you a heads up
Thanks, I need to get off Zight, they used to be such an nice option for fast file share but they've really suffered some of the worst enshittification I've seen yet.
Links are broken.
So.. sign up. "Get Sight for free". Ads everywhere bro.