does anyone know or have an opinion which image style or image model (or combination) is least likely to invent words or use text in images? I've attempted to include 'do not include text' in image prompts, but im ignored like im back in high school - looking for the best results that don't involve editing since I'm using the api to send generations directyl to users. my users have the ability to select an image style, but want to recommend a good default for best results.
when you hover it has details about the model and some explicitly mention good for text. I like Ideogram 3, Flux Kontext and Flux Context Pro.
Mariena Q. is there one that's good for not adding text? or less likely to? would love to get images free of text. the generations go to real estate agents or mortgage lenders and every now and then there will be an image that will include a made up name for a company.
Is it adding a “for sale” sign? I wonder if you can test different language. If you think about the data the model is trained with real estate agents are generally associated with signs with text in front of a house. So maybe say a homeowner in business casual?
it does produce 'for sale signs' and 'sold signs', but those are totally fine. not all of the text it creates is bad (it will have 'contracts' or 'closing' or 'valuation' on docs - a lot of relevant text, some phrases like 'your home journey' or similar are relevant and looks great. there are just regular instances where it will come up with a fake brokerage or lender name in the image like 'Apex Realty' or 'Evergreen Realty' - stuff like that. then in some instances with photo realistic options it will have some misspelled or jibberish words. easy to edit, but most of my users will not be able to do that and in many instances generations will be automated/going direct to their clients. it doesn't happen all the time, 30% with some styles, 10% with others, but would be ideal to minimize it. 'no images' is an option of course, but the generations just look so much better with them. I know it's a common problem in AI. apparently Will Smith found out the hard way this past week.
Yeah I would keep a human in the loop to do editing before it goes to a client for that reason. I’ve not heard of a one-shot solution
.png)