The latest algorithm for Generative Fill seems suddenly dumber?

[Jen27654846oo5f]

I've been using Generative Fill a few months to change photos.  In the last week, it seems to have gotten dumber.  One example is to select a head and enter "cowboy hat."  Before, it wasn't perfect, but everything that came out was essentially a cowboy hat.  Now, I'm getting so much content that is clearly not a cowboy hat - baseball caps, beanies, berets?!  A cowboy hat appears maybe 20% of the time, and worse quality.  This is disappointing, I hope you can fix this because it's such a cool feature when it works.

[Pete.Green]

Hi @Jen27654846oo5f ,

Are you using the beta or the release Photoshop?

With beta, you can also use Reference Image to point to the type of hat you're after as well. 

If either one, you can also use the thumbs up and down to report generations as positive or negative in quality or in result. That's great feedback for the team as they work to improve these generative models.

Regards,

Pete

[John Metzger]

I would echo what Pete said about leaning into the use of a reference image (works best with a clear subject on a white or black background). By providing a bit of additional input into the model you should get much better results. 

This verison of the model was trained to give more variability in the output but higher prompt adherence (you should be getting cowboy hats but more variety in them). Clearly we have further to go in getting to that objective but use of the reference image should get you back on track. 

Btw - if you are interested in sending me the files (PSD including the masks and prompts and generative layers created) that are proving to generate problematic results. I'm happy to bring them to the attention of the people building out the latest model so they can investigate why the results are poor.