Generative Fill in Photoshop Beta

I ran into this trying to use generative fill on some old industrial photos I had taken of a now demolished coke plant. Coke is a coal product made from heating coal without oxygen, similer to how charcoal is made from wood. It has no relation to the soft drink and the images were not of a Cocacola company facility. I started out being specific with my prompts and included this term, which resulted in a violation with each generation. The prompts all included coal, industrial, plant, etc. so it was not without context. It was not clear how or why I was violating guidlines. No guidence was given. When I began to suspect the word "coke" might be the culprit I eliminated it from the prompt and had no more issues.

This was a niche scenario, but the fact that real words are being blocked accross the board because they are also used by big brands is a serious issue. I can understand wanting to filter against recrating brand logos, but this should not be done by filtering out brand associated words regardless of context. The fact that this happens without any guidence as to what caused the "violation" makes it an even deeper issue.

On strange thing in generative fill is if I want to increase crop the image and add a missing arm or leg I get the not allowed message. I found a work around, just use the word suntan. Sometimes it gives a good result on the first try, then again it may take many. Sometimes it will only put a large sun hat in place of the limb. From distored grossly under or oversized to believable. Sometimes no matter how many times you try the results in bad, especially with feet. It wanted to put the sole up many times before it was right. 

Hello, I'm here to share ideas and suggestions for GenFill.

I work on complex psd files with lots of photos and layers, like sometimes 50/80 layers. I don't start from zero or from a simple image, I prepare and do a lot of images research and work on complex photomontages for video game background. All this to say that I used several sewing methods, including aware content, and now I also use GenFill in my work. And there's a slight problem with using it on psds when you're trying to do what I do. Let me explain.

To begin with, I'd like to remind you that content aware works on the specific layer selected. And that until now, GenFill works globally on the image, not on the layers. I find the 2 tasks interesting, depending of what is for. I've become accustomed to aware content and find that the fact that it works on a specified layer makes it quite noticeable and easy to use. The only thing is that it doesn't allow you to manage or generate randomized patterns. Improving it would reduce your genfill server load.

The fact that GenFill works on the overall image poses a number of problems in terms of quality of life.

The first thing is that it doesn't allow you to direct correctly on specific intentions. If, for example, you want to work on the background mountain only, which is on a specific layer, it will necessarily make the join of his choice not your with the rest. This obviously means disabling all layers to use only the specific object.

A second constraining factor is when certain layer adjustments such as color, level etc. are used.
The problem is that genfill takes a layer adjustment into consideration when it shouldn't, so it ends up applying it twice, once to the generated image + the layer adjustment that's still there. This generates serious chromatic problems, forcing all layer adjustments to be deactivated.

You'll tell me that it would be enough to display only the layer by alt-clicking on it and then alt-clicking again to display all the layers. But as genfill has generated a new layer, this handy function is not usable here. But I'm not a Photoshop expert, so maybe another solution exists.

In any case, the 2 combined are quite problematic. They make life difficult to use with many layers and when you're in the middle of work. An option to choose global image or specific layer would make things a lot easier.

Thanks for reading.

In an article on Photoshop (Beta) Generative AI I've just completd,  I provide seven Gen AI examples with prompt - all based on one photograph I took of F1 cars racing at the Canadian Formula1 Grand Prix in June and I analyze how effectively Gen AI interpreted the prompt. 

This is the original image I shot of the F1 cars coming out of the short straight just before the S curve where we were seated at the Canadian Formula 1 Grand Prix in Montréal in June. Two of the seven (total) AI Generative Fill images below are derived from this one. 

Here’s the prompt I used: “make cars appear to drive on water going through tropical forest.” I’m impressed! Gen AI did a great job suggesting reality in an unreal situation, the displacement and splashing of the water create a convincing sense of movement and speed in an entirely different medium of water. The color, leaves and contour on both sides of the forest look realistic.  In addition, the image shows one of Gen AI’s strengths - its ability to render nature well. Note the perspective, lighting and scale - excellent. A+

The prompt was to make it appear as if the cars were out in space driving through the solar system and the Starship Enterprise was in the background  - but I don’t have it verbatim.  I think I see the Starship Enterprise in the mid- right border but I think it’s a fun, fantasy of space, although the cars look as if they are floating rather than racing.  I suspect others have discovered a Must Fix in Adobe’s  next version that doesn’t exist in Beta - find a way to save the prompt with the image.  B+

I'm writing a book for authors using AI. Here is a pdf showing how I'm explaining Photoshop's generative infill in this book

[personal information removed as per forum guidelines]

https://community.adobe.com/t5/using-the-community-discussions/adobe-support-community-guidelines/td-p/4788157

Hiya folks here's a few pics I've created and spent hours tweaking. Worth the effort though.

Maybe admin could move to the correct board if I'm in the wrong place ?

I wanted to add a "zepplin" in the sky and that gave a violation of guidlines. Seems like I mispelled the word (it should have been "zeppelin" but then I got a plane...). AI should regognise that it is a spelling mistake and either generate what asked for or suggest the correct spelling (like MS Word does e.g.) 😕😕 Seems like Adobe really needs to work on this (generate the right image + allowing/correcting spelling mistakes) if they want to catch up with AI...