Photoshop AI and Generative Fill and the Flaw It Creates
ฝัง
- เผยแพร่เมื่อ 29 มิ.ย. 2024
- In this video and quick look at one flow of the Generative Fill feature of Photoshop and how to spot it.
Subscribe for editing tips and behind-the-scenes videos as well as community contests and free giveaways:
www.patreon.com/WildlifeInspired
-------------------------------------------------------------------------
Best noise reduction and sharpening software: topazlabs.com/ref/351/
-------------------------------------------------------------------------
MT FAVORITE GEAR
Nikon Z9: bhpho.to/3rM9O7A
Nikon 400mm 2.8 Z: bhpho.to/3tj806y
Nikon 105mm macro: bhpho.to/3NngzVI
Laowa 25mm ultra macro: bhpho.to/3F405g7
Leofoto monopod: bit.ly/46xfQYE
Robus Tripod: bhpho.to/46eAQn3
Promedia Gear Tomahawk (monopod gimbal): tinyurl.com/4vdkuy98
Promedia Katana Jr (tripod gimbal): tinyurl.com/yjtf4s4a
Leofoto fluid head: bit.ly/3ZFtRB8
AK Diffuser: akdiffuser.com/?wpam_id=30 (COUPON CODE skeys5 for more savings)
-------------------------------------------------------------------------
Shopping for any of these companies? Use these links:
B&H (my online source with great delivery and service): bhpho.to/41vuugj
Promedia Gear: www.promediagear.com?aff=36
Leofoto: leofotousa.com/wildlifeinspired
As someone who also works frequently on commercial style composites with wildlife being a side hobby Gen AI has become a huge part of my life.
What I've noticed with PS's gen AI is the following:
1. It was trained on Adobe stock which means it is great at creating stuff that is commonly found in stock images but it is awful at creating more imaginative stuff. (Compare to something like midjourney which was trained on creative work so is great at being imaginative)
2. The "mask" it creates often is a terrible blend. You can instantly improve your results by taking some time and blending the mask yourself.
3. As you mention, it can't match noise profile very well. I try to work with as little noise as possible to begin with but I find often introducing a new noise profile across the whole image as a final step can hide things quite well.
4. The resolution sucks, which often doesn't matter, if you are generating a small area or an out of focus area you won't even notice the fact that your generated tile is a 1024x1024 image stretched to whatever area you are filling but if you are working with a detailed part of the image it can be painfully obvious. I often have to address this by pulling the generated area out of Photoshop and running it through an upscaler like Magnific AI before bringing it back to Photoshop to work with.
Gen AI is progressing super quickly, scary quickly, but it still has some massive limitations. I'd expect that in 12-18 months though the act of using photoshop will just be about executing creative vision as the actual craftsmanship of image editing will be irrelevant. (This is mostly a good thing)
There are many times that "Content Aware Fill" remains a better choice over "Generative Fill". I think you brought up a great point and it is worth a future video for when in the processing flow, should Generative Fill and Topaz Denoise be used for the best results.
I've used the expand feature to complete the wings of birds that I've ended up clipping as they took off. I'll only do it if I otherwise like the image. It's a real grab bag though as far as quality goes.
Same can be said of the new Remove Tool. Excellent when dealing with small, or thin areas, or spots with lots of details, but leaves behind weird textures in smooth and wider areas when inspecting closer. Great at what it does, in moderation. Looking forward to improvements as new upgrades become available!
I have used Gen expand with good results but there are some fixes like you mention. The Gen AI are low res so fixing noise might help but some patching also works but time consuming. Also, software such as Magnific AI helps enhance the generated part. Unfortunately, it is costly unless one uses it a lot.
Perhaps you were just going quick & loose with the lasso tool for the sake of demonstration, but I find that I get much better results if I draw much closer to the edge of the subject that you want to delete. It may make no difference though with the difference in noise pattern that you showed.
Great video. I'm surprised to see that on the first photo (bird feeder) you could plainly see a line where the lasso was, even without any special inspection settings.
Yeah, I’ve never seen it perform that poorly on such a simple task. I don’t think that example was representative of its potential. But I think Scott does still have a lot of good tips.
I watched Ray Henesy deal with this issue of noise patterns by reintroducing it after the fill.
Generative fill is just inpainting using adobe's (really bad) firefly model. A photographer can still inpaint/outpaint with stable diffusion and midjourney and you would never know. For one, adobe will inpaint just the masked area, but with stable diffusion, you can tell it to regenerate the whole image, or do 'soft inpainting', and it'll look worlds better.
I know this is a results of the feature kind of video, not an ethics video, but I refuse to use modern photoshop/cameras for this reason.