Comment by deathanatos
7 days ago
… sure … but also no. For example, say I have an image. 3 people in it; there is a speech bubble above the person on the right that reads "I'A'T AY RO HERT YOU THE SAP!"¹
I give it,
Reposition the text bubble to be coming from the middle character.
DO NOT modify the poses or features of the actual characters.
Now sure, specs are hard. Gemini removed the text bubble entirely. Whatever, let's just try again:
Place a speech bubble on the image. The "tail" of the bubble should make it appear that the middle (red-headed) girl is talking. The speech bubble should read "Hide the vodka." Use a Comic Sans like font. DO NOT place the bubble on the right.
DO NOT modify the characters in the image.
There's only one red-head in the image; she's the middle character. We get a speech bubble, correctly positioned, but with a sans-serif, Arial-ish font, not Comic Sans. It reads "Hide the vokda" (sic). The facial expression of the middle character has changed.
Yes, specs are hard. Defining a spec is hard. But Gemini struggles to follow the specification given. Whole sessions are like this, and absolute struggle to get basic directions followed.
You can even see here that I & the author have started to learn the SHOUT AT IT rule. I suppose I should try more bulleted lists. Someone might learn, through experimentation "okay, the AI has these hidden idiosyncrasies that I can abuse to get what I want" but … that's not a good thing, that's just an undocumented API with a terrible UX.
(¹because that is what the AI on a previous step generated. No, that's not what was asked for. I am astounded TFA generated an NYT logo for this reason.)
You're right, of course. These models have deficiencies in their understanding related to the sophistication of the text encoder and it's relationship to the underlying tokenizer.
Which is exactly why the current discourse is about 'who does it best' (IMO, the flux series is top dog here. No one else currently strikes the proper balance between following style / composition / text rendering quite as well). That said, even flux is pretty tricky to prompt - it's really, really easy to step on your own toes here - for example, by giving conflicting(ish) prompts "The scene is shot from a high angle. We see the bottom of a passenger jet".
Talking to designers has the same problem. "I want a nice, clean logo of a distressed dog head. It should be sharp with a gritty feel". For the person defining the spec, they actually do have a vision that fits each criteria in some way, but it's unclear which parts apply to what.
The NYT logo being rendered well makes sense because it's a logo, not a textual concept.