Generative Fill produces blurry results

[Macro Palermo - MRC Foto]

Hello everyone, I'm trying out the new generative fill feature that uses artificial intelligence, and it's truly impressive! However, I'm currently experiencing an issue, and I'm not sure if it requires specific settings or if it's a problem that still needs to be resolved since it's in beta. The problem I'm facing is that when I generate an AI-assisted edit, the resulting image is blurry and unusable. I'm attaching an example here. In this case, I enlarged the original image and had the AI reconstruct parts of the background that didn't exist. If you zoom in, you'll notice the clear difference between the original photo and the reconstructed part. Am I doing something wrong with the settings? Please let me know. Thank you.

I am using a Macbook Pro M1 Max

Mac Os 13.3.1 (a) (22E772610a)

[asfandy38962123 · Accepted solution]

AI is limited to resolution of 1024x1024, Meaning it will not generate clear results if your image is bigger than that (There is a workaround thought)

Please see this video: https://www.youtube.com/watch?v=pGJ0KjT4Xds

This will fix your issue most likely 🙂

[c.pfaffenbichler]

I wonder how this will be addressed if and when the feature gets a general release; the transmission of the data for images larger than 1024px x 1024px to and from the Adobe servers seems like a possibly lasting bottleneck … 

[AKS666]

So today I installed version 25 of Photoshop with generative fill officially open for commercial use.

Still the generated images are blurry and of low resolution. Not really of any use...

So when is there gonna be high resolution support coming?

[A.Britt]

The much advertised Photoshop AI is simply bad. The results are unusable for me.

[Trevor.Dennis]

The much advertised Photoshop AI is simply bad. The results are unusable for me.

By @A.Britt

We can help you fix that if you like.  Tell us where you are having difficulty, and we'll show how.

In fact, if you just check out The Photoshop Training Channel, and Photoshop Cafe YouTube channels, and search them for Generative Fill, I think you'll be golden.

[A.Britt]

Could you provide a specific site? If it is a work-around, it is still dissapointing..  I've been with PS for decades. The AI is not up to Adobe standards.

[c.pfaffenbichler]

You need not use the feature if it does not meet your expectations. 

Otherwise you can provide feeback to help improve it, but whether it will ever live up to your standards seems unclear. 

[default2rkbo39it64o]

I Agree, its really bad, at first glance some of the results can seem impressive, but just zoom in a little and it is clearly a very bad low resolution solution that doesn't blen well with the original image. Absolutley no use to print designers as it stands. Lets hope that Adobe either drop the price of the subscription to match the poor quality or wait until they have improved it and then releas a better version. 

[default2rkbo39it64o]

I have noticed the same problem, its just bad. If we are all going to have to go back and fix the problems that AI have created, then really, what is the point in the first place. 

[default2rkbo39it64o]

I agree and have noticed the same problem, AI results are very poor and not fit for purpose. I have noticed that when anyone, including me makes comments about it, community experts are very quick to try and shut down these points of view. It is very clear and obvious that AI is going to degrade the quality of creatives work and ultimately drive the price creatives can charge down with it, meanwhile Adobe will continue to hike up its price every year in return for products that are getting worse due to the inclusion of AI. Please can adobe address this problem and make it very clear to everyone who has or is thinking of having an Adobe subscription that it is not for professional use and that professional creatives work at a much higher standard than the AI results.

[c.pfaffenbichler]

Are you (aside from monetary issues and community experts disagreeing with you) complaining about the implementation of Generative AI in Photoshop at all or just the quality of its output (edit: or both)? 

Please can adobe address this problem and make it very clear to everyone who has or is thinking of having an Adobe subscription that it is not for professional use and that professional creatives work at a much higher standard than the AI results.

Professionals (not all, but at least some) use Photoshop’s Generative AI-related features (which includes Remove Tool etc.). 

If they remember to check the results and, if necessary, just use the techniques they have been using previously in digital image editing to amend or completely redo the created content why would they not? 

[default2rkbo39it64o]

Both, AI is fundimentally wrong for humanity, why invent a tool to replace what humans do? I really can't understand this. As it stands, adobes use of AI is very poor and anyone can see that, what is the point of a tool that does the job worse than a talented creative just for the sake of it. I agree with you that some professionals might try AI to see if it helps, but when you are having to completely redo what AI has produced then how is that of any benefit? Generative fill is an absolute joke, AI can't even recognise what a human being should look like, I have tried it and if you zoom in on the faces or the hands you get some weird frankenstein like creation that doesn't pass for a human being. There is no attempt by Adobe to explain that AI can't be used in print design because of the resolution problems, so people who don't know enough about its limitations are going to think that print designers can do the work in half the time and not pay them as much. AI devalues the work of creatives, it has created a false idea that things can be done quicker and just as well as traditional methods, but it is just not true. Maybe a quick visual for a social media post might be a use, but even then, just take an extra bit of time to curate your own assets and carefully comp it together, it will look much better every single time.

[c.pfaffenbichler]

That Adobe’s Firefly produces worse results for people (and other animals) than some of the competitors’ software seems undeniable. And whether they will ever catch up remains to be seen. 

I recently had to add missing content for blurred/»out-of-focus« backgrounds (without people) behind product shots – for such stuff Generative Fill seems just fine. 

And I probably use the Remove Tool more than the Healing Brush Tool now, a more modest implementation but it still utilizes AI. 

As for AI more generally:

There are not just ChatGPT, Gemini, Stable Diffusion, Midjourney, Firefly etc. but also stuff like AlphaFold, which has supposedly pretty much solved the prediction of Protein folding, which seems like quite an impressive achievement. 

[Dylan27385084emhi]

Hello. 

I don't understand. Is it a Photoshop problem or is the procedure I'm using incorrect? I followed the video perfectly but the generated images are blurry and of low resolution.

[c.pfaffenbichler]

Please provide meaningful information; what are the image’s pixel dimensions, what were the Selections pixel dimensions, …? 

Generative Fill creates a 2000px x 2000px images and upsamples that if necessary. 

[Dylan27385084emhi]

the image has a resolution of approximately 3450x2800. I increased the width of the image by 1014px and the selection measured 1024px x 1024px. I also tried with 900px solutions but the result was the same. 

[c.pfaffenbichler]

Could you please post screenshots taken at View > 100% with the pertinent Panels (Toolbar, Layers, Info, Options Bar, …) visible? 

[Dylan27385084emhi]

of course. the image has a resolution of approximately 3450x1450. I post the screenshots.

[c.pfaffenbichler]

I seem to get a (slightly) better result with a small Selection. 

[defaultn7mbfckh78c6]

Thank you for posting this. It is happening to me as well and apparently to countless other people as a simple search "Generative Fill Photoshop blurry images" would demostrate. I use large images 6000x4000 8bit 300 dpi  RGB mode. Usually I extend using the crop tool with Generative Fill to about 8400x5600 and I experience the same atrocious results.  On screen at 100% seems to be fine but once printed at about 1 meters away from the print one can clearly see the distinction between AI generated vs original photo.  Although a pretty good tool for amateurish results, I guess we are not there yet. I am afraid Adobe may be victim of degradation of the original software development "philosophy" RERO (Release Early Release Often) or Continuous Deployment or Rapid Deployment, which it states if it's good enough release it. Don't forget they need to keep share holders interested and prove that they have not missed the AI bandwagon.

[c.pfaffenbichler]

If you expand the canvas in one go you will get a 2000px x 2000px image upsampled to the final size (8400px x 5600px). 

One can work-around that by extending in smaller sections (smaller than 2000px in either direction). 

Of course this can be cumbersome and whether the result meets one’s expectations is another question. 

[teodor3338480027do]

of course its cumbersome and not a valid workflow especially when you need to edit hundreds of pictures.

theres nothing subjective about expecting a product to do what its advertised to do, its just false advertising for a crappy product.

[c.pfaffenbichler]

The Generative Fill-results for Selections above 1024px x 1024px can certainly show huge quality differences to the original image content, but if you want to make an accusation of »false advertising« please provide a Link to a page where some Adobe representative lied about those limitations or made other counterfactual statements regarding the feature. 

[teodor3338480027do]

yea, its the consummers fault for imagining the possibilities

its this trend of over-promising (and selling products on their potential not the actual value) that seems to be the norm today in every market