Transcription
Welcome to another video tutorial. Today I'll show you how to refine and enhance your AI generated images from ChatgPT or other similar platforms using Photoshop. So stick around until the end to pick up a few tricks.
I use Chat GPT40 to generate this landscape image. As you know, one annoying thing about ChatGpt is that it can't do 16 to9 ratios, only 3:2. So let's copy the image or download it with this button.
In Photoshop, when you create a new document, it automatically picks up the image size from the clipboard. All ChatGpt landscape images share these dimensions. Now I can paste the image. The problem is I want it wider so I can use it on YouTube and other platforms without black bars on the sides.
If I grab the crop tool, you'll see I can crop to different sizes. And the top right corner shows the dimensions as I move the crop box. You can eyeball it, but what if you need an exact ratio? Say I want a square crop. How do you set that? In the ratio dropdown, pick one of the suggested ratios or make your own preset. With the square preset, the crop box stays perfectly square no matter how you drag it. It's handy that Photoshop shows the rule of thirds overlay, too, so you can align your subject along those lines. Press enter or click the check mark to apply the crop.
What I like to do next is click the little arrow next to the preset menu and switch to document dimensions. That way I can see the exact pixel size of the cropped image. Let's select the crop tool again and this time I'll choose the 16 to9 ratio. You can also enter any ratio manually. Then drag from the corners to adjust the size. Normally you'll keep the full height and expand on one or both sides depending on your composition, but that leaves you with a transparent border which I don't want. So undo a few steps.
Reselect the crop tool with the same ratio and under fill, choose content aare. If your edges aren't too complex, content aware fill does a decent job. There's just a tiny repeating pattern here and a slight blur line there, but it's barely noticeable. Undo again, and let me show the method I prefer. Crop your image as before. Then under fill, choose generative expand. This uses Adobe's cloud AI to seamlessly extend the edges. Note that NSFW content won't expand and each expand uses credits from your Adobe account. You might want to keep an eye on your balance. Here's the final result. A perfect 16 to9 image. Make sure the properties panel is open so you can review the three AI generated variations. If you're not happy, you can rerun the expand. I like the middle version best, so I'll keep that. Finally, open the layer menu, choose flatten image, and then save your new 16-9 image as a JPG or another format.
Here's another example using a portrait ratio version of this futuristic image. Those are the dimensions Chat GPT uses for all portrait images. I just pasted the image. Copying and pasting is faster for me than saving and loading. I grab the crop tool again, then click the arrows to switch from 16 to 9 over to 9 to 16. Since there's no important detail at the sides, you can crop it straight to portrait without needing generative expand. But if you do need more space, crop as usual, choose generative expand and wait. Depending on the image, it can take around 15 seconds. Here, there wasn't much detail at the top beyond sky. So contentaware fill would probably work even quicker on a simple background. You can also rightclick any layer and choose flatten image from there. Now it's almost full HD. To make it exact, go to image image size, turn on reample, and enter the full HD dimensions. That makes it perfect for shorts, Instagram, Tik Tok, and so on. You can pick one of the enlargement presets or leave it on auto. And here's the final image.
And one more example. I have this nice photo of macarons, but you can see it's cropped at the bottom and I want the full image. I could ask ChatPT for another version, but it would give me a different picture. So, I prefer Photoshop to fix and even enlarge it. I grab the crop tool and set it to a square ratio, then expand the crop to leave enough space at the bottom. If I try contentaware fill, it fails on complex backgrounds as you can see here. So, I undo, recrop with the same ratio, and this time choose generative expand for the fill. Look at that. How amazing is that? Then I just picked the best of the three AI generated options. It's really useful.
I generated a refreshing drink photo in a square ratio, which gives you a 1,024 pixel image. So, how can we make it bigger? One solution is to open the image menu and choose image size. Then make sure reample is checked. So, Photoshop adds pixels. Keep in mind this can reduce quality by altering the original pixels. Let's say I double the size and choose whether to smooth it more or preserve some detail. When I click okay, I get a larger image, but the quality is only okay, not great. So, let's undo that and try another method.
Go to the filter menu and select neural filters. On the right, you'll see several options. Enable the super zoom filter. For output, pick new document so your upscaled image opens in its own file. Then click the zoom icon. It will show 2x and update the dimensions, but the change isn't applied until you hit okay. You can push it higher, but the more you enlarge, the more the quality suffers. You can also click the show original icon to compare before and after. I'll stick with a 2x zoom. You can tweak the noise reduction sliders. Use them subtly. I'm cranking them up so you can see the effect. and adjust sharpness up to 100 for a true size zoom. I'll add just a touch of sharpness since the image is a bit soft. That gives you an upscaled image. You always lose a bit of quality when enlarging, but this method delivers a cleaner result when you need a bigger version.
So, are there any better upscalers? Let's take the original image again, and this time I'll test it on a specialized upscaler. I'm going to use Topaz Gigapixel AI, which is a premium software with more options when it comes to upscaling and uses AI to do it. You can choose how much you want to upscale, just like in Photoshop, but you also have the option to select different models depending on the image. Some models are better for photos, others for graphics, and so on. You can click on the image to see the before and after, but the higher you go, the more the quality will suffer. It's still better than just enlarging the image since it re-creates the pixels, but you'll never get the same quality as a photo are originally taken at that large size. Once you're happy with the settings, you can export the image. You can also do batch upscaling and other tasks. If you're thinking about getting the software, try waiting for Black Friday or other holidays since they usually offer discounts. It's quite expensive at full price. For this image, there wasn't a huge difference between Photoshop and Topaz Gigapixel. Photoshop had more artifacts, and Topaz reduced some of the noise and artifacts, but I didn't mess with the sliders too much. So, try different options and see what works best for you.
By the way, when you use Super Zoom, it also changes the resolution from 72 PPI to 300 PPI during the upscale. Topaz also lets you choose the PPI you want. Let's switch back to the original image, which has only 72 ppi and 1024 pixels. Let's say I don't want to mess with the pixel count. I just want to change the print size and have it at 300 ppi. To do that, you need to go to image size. And if I try to change the resolution to 300 right now, you'll see the number of pixels changes. And I don't want that. I want the same pixel dimensions, just a different print size. So, what you need to do is uncheck resample. Now, it shows the size in centimeters. Uh, at 300 ppi, the print size is only a few centimeters. The pixel count stays the same, but we've converted the resolution from 72 to 300 ppi. If we go back and change it again to 72 ppi, you'll see the print size becomes much larger, but the resolution isn't great. So, now we're back to the original image we got from ChatgPT with those settings. Just remember, if you want to change the pixel size, check resample. If you want to keep the same pixel size and only change the print resolution, uncheck reample.
Let's move to the next use case, mockups. I generated this white mug photo, which turned out pretty nice. Then I created an illustration just so we could test something. I asked for it to be on a transparent PNG background to make things easier to work with. In this chat, I asked chat GPT to generate a photo of the illustration on the white mug, and I got this image. At first glance, it looks like it did a great job. But did it? Let's take a closer look. Here's the original mug image. And here's the result. As you can see, the mug, the color, and even the position are different. It's not my image of the mug. It's a completely different one. But I can live with that since a mug is still a mug, right? But what about the illustration? Here's the original illustration. And here's the one on the mug. The image is different. The woman's face, the cat's eyes, they've changed. Now, that can be a real problem if you're doing mock-ups with your illustrations, labels, or logos. You don't want AI to change the design. You want it to be identical. So, I use Photoshop to handle that.
First, we'll keep only the photo of the mug. Then, using the rectangle tool with shape mode selected, I draw a square or rectangle that will fit nicely on the mug. Since most mug design usually have that kind of ratio, a size like this should work. I'll rename the shape layer to design, then rightclick and convert it to a smart object. Now, the shape of the mug is a little round. So, let's adjust the square to fit better. I'll go to edit transform and choose warp. Or you can just use free transform, right click, and pick the option you want from there, like warp. Warp lets you move the points however you want, so you can shape it freely. But let's say I go with an easier method. I'll undo that and instead of using custom, I'll choose arch. There's a newer option called cylinder, which might work even better since the mug is a cylinder, but let's stick with arch for now. You'll get a handle you can drag to control how strong the arch is. Let's say I apply just a slight arch so the top curve matches the mug a bit better. Then press the check mark button to apply it. Now the warp is applied to the smart object which acts like a container that holds your design. If I double click the thumbnail, a new document will open in PSB format, not PSD. And here you can add layers or your design just like in any normal document. Let's say I hide the square and now add my illustration. Once it's added, I close the document and it will ask me if I want to save. Choose yes. And now that smart object is updated. You'll see the update both in the thumbnail and on the mug itself, but the lighting and shadows don't quite look realistic yet. So, let's see how we can improve that.
I'll duplicate the mug image and move it on top of the illustration. Then, I'll hold the Alt key and hover between the two layers. You'll see the cursor change to a square with a down arrow. If I click between the layers, it creates a clipping mask. So, now the mug image is only visible inside the area of the illustration. Next, I changed the blending mode of the mug image to multiply. A lot of people just stop there. Changing it to multiply is common when the mockup has a white background because it helps show the shadows. But I like to take it a bit further. I reduce the opacity of the multiply layer depending on the design and I rename that layer to shadows. Then I duplicate that layer and name the new one light. For this one, I change the blending mode to screen so I can add some light as well. just a tiny bit. Nothing too strong. You can play with both the shadow and light layers until the design starts to look natural to you. Now it's looking pretty good. I'll save the file as a PSD so I can use it later. And I'll also try to upload it to Discord along with the prompts so you can download everything.
Now, you might be thinking that's a lot of work just to make a mockup. But here's the cool part. Now that the mockup is done, you can just doubleclick the smart object, hide or delete the current illustration, and add a new one. Resize it however you like. Close the smart object, save it, and boom, the new illustration appears on the mug. This way, you get your exact design, not a changed or randomly generated one like chat GPT might give you. And depending on the colors, you can tweak the shadows and light until it looks just right. You can add all kinds of images or text inside that smart object and make the mockup look exactly how you want.
Moving on to the next use case. I have this portrait of a woman and I want to remove the background. So, I asked chat GPT to generate the image again, but this time with a transparent PNG background. I got an image with a transparent background, which is good. Now, if you're using AI generated images, it's usually fine if the new version is slightly different. But if you need the exact same face or design, then getting a different version is not okay. If I compare the original image with the AI version that has the transparent background, you can see they're similar, but they're not the same. So, in cases like this, I prefer using Photoshop to remove the background myself.
If you've updated Photoshop recently, go to preferences, image processing, and under select subject and remove background, choose cloud. It's a bit slower, but it does a much better job at cutting out the background. Click okay and restart Photoshop. Now you should see a new bar called the contextual taskbar. If you don't see it, go to the window menu and enable it by selecting contextual taskbar. Watch what happens when I click remove background. Look how cleanly it cuts out the background. Same face, same image, just the background removed exactly how I wanted. And that's not all. You can now click generate background. Let's say I want something exotic like an island with turquoise water. Just hit generate and wait for the AI to finish. You'll get three different backgrounds. If you need more variations, just click generate again. Look how perfectly it replaces the background. You can't even tell it wasn't the original.
So, we saw how to remove the background, but what if we want to remove something from the image itself? Let's test it in Photoshop. If you select the remove tool, the latest version of Photoshop has this feature called find distractions. It lets you remove things like wires, cables, or even people with one click. I've got these cables in the image. Let's see how Photoshop handles them. It removed all the cables from the scene. Some of them had vegetation over them and it left parts of that floating in the air, but I can just paint over those with the remove tool and when I release the mouse, it cleans them up. I use the remove tool for all kinds of things to get rid of imperfections, clutter, or unwanted elements. I usually keep generative AI turned off by default because it runs faster, but if you're dealing with more complex removals, you can set it to auto or turn it on.
Now, what if I do I want to remove this dog? I can just paint over it, maybe with a slightly larger selection to include the shadow, and it does a pretty good job cleaning it up. If I want to remove the person, I can do that, too. And then use generative fill to place something else in the same spot. You can quickly select the subject from the toolbar. Once selected, I can press Ctrl + J to jump that selection onto a new layer. That way, I can move it and duplicate the person. Um, let's say I want to have three people in total. I'll flatten the image, select all, copy the image, and paste it into chat GPT and explain what changes I want. Chad GPT can understand much better when I send it an image. So, I use this method all the time for quick manipulations in Photoshop. I show Chad GPT what I want and let it generate the result.
While Chadpt is generating the image, I can also try doing the same thing in Photoshop. Generate three different people so they don't all look the same. Looks like Photoshop did a good job, too. It looks pretty convincing. Chad GPT also followed my suggestions. I now have three people, no cables in the scene, and the lighting looks like it's during golden hour. Now, let's take that image and paste it back into Photoshop to continue editing. I'll take the lasso tool and make a selection on the right side of the image. Then, I'll use generative fill to generate a dinosaur. Photoshop gave me three versions, and I think I'll go with this one.
Over in chat GPT, there's also an inpaint feature. You can use the brush tool to make a selection and then describe what you want to appear there. Now, it's generating a new image, but for some reason, this time it added the dinosaur, just not where the selection was. Um, pretty strange. Usually, it works fine. I pasted the image that Photoshop gave me into chat GPT to try to guide it better. It improved a bit and got closer to what I wanted, but the dinosaur still wasn't in the right position. Also, it changed the look of the image. It smoothed it out more than I wanted. So in case like this, Photoshop is more reliable. It can help fix or handle things better when chat GPT doesn't quite get it right.
For example, with this photo of a lighthouse, I'll go to filter and choose camera raw filter. Here you have all kinds of options. You can adjust the temperature, play with the exposure, move the sliders to see what they do, and then dial them back to apply subtle changes so you don't overedit the image. In this case, adjusting the shadows can help remove some of the darkness. You can also try the whites or blacks sliders to see if they help. I'm just showing what options you have. I'm not trying to fully edit the image right now. Tint can also be useful depending on the image. With saturation, you can even make the photo black and white, but you have better tools for that under the black and white tab. From the effects tab, you can increase texture or add a soft, blurry glow. Clarity is sometimes useful, too, especially if you want that HDR style look. Dehazy is great for foggy photos, but it can also be used creatively to get a fantasy style effect or to add mood. So, you can use Camera Raw, not just for basic corrections, but also for creative manipulations. There are tons of tutorials on YouTube focused on photo editing. It's a really useful skill to have. The lens blur option is also interesting. You can adjust the blur amount, and the cool part is you can move the focus point. This lets you create some unique and dramatic images.
With the object selection tool, you can select individual objects to color correct or edit only those parts. You can also make a selection around multiple objects if needed. And don't forget about selective color. It's great for subtle adjustments to specific colors without affecting the whole image. Photoshop is a complex tool, but also a very powerful one. From the select menu, you can quickly select the subject, the sky, or other elements and apply masks to target just those areas.
I also create a lot of icon sets with chat GPT. I usually ask for different icons or multiple variations of the same icon. Then I pick one that I like and sometimes regenerate it for better results. The image usually comes at 10 24 pixels. So if I divide that by two, since each icon takes up half the space, I get a 512 pixel icon. So I go to image canvas size and set the width and height to 512. This works if the background is either not locked or on a separate layer like I have here. Now I can move the image so that only one icon is visible and I export it as JPG or PNG. Then I move to the next icon, position it and export again, repeating that for each icon. Another thing you can do is edit the colors, paint over parts, or change it however you want.
Once you're happy, you can flatten the image, copy it, and paste it back into chat GPT, asking it to generate the same image, but in higher quality. Now I get a similar version, but at 1024 pixels again. I bring that back into Photoshop to make subtle adjustments. I can also remove the background or replace it entirely. In this case, I'll remove the background and apply a layer mask. Next, I go to image trim and choose the option based on transparent pixels. When I click okay, it crops the image down to the exact size of the visible pixels. Super useful for clean exports. Then I quickly jump into Camera Raw to make a few small adjustments because it looks a bit too dark for my taste. Maybe I'll tweak the colors, too, just to make them more vibrant.
So, yeah, I hope you learned something from today's video. There's so many things you can do with this workflow. Here's how to get the prompts and resources from this video for free. At the top of my YouTube channel page or in the video description, you'll find a Discord link. Click it, then click go to site and accept the invite. You'll land in the welcome channel. The easiest way to find what you need is either to click the help channel in the sidebar or to type hash symbol and then help in the message box and hit enter. That will create a link. Click it to jump to the help channel. Once there, look for the AI2Play section. That's where all the resources and prompts live. Click on it and you'll see prompts and images for today's video on these seven practical use cases. You can scroll through to find everything. We have over 27,000 members and plenty of channels to explore. If you ever get lost, just use the help channel to get directly where you need to go.
That's all for today. I want to thank everyone who subscribed to my membership and supports this channel so I can keep making tutorials. A special thank you to AI Titans, my highest level supporters. Don't forget, if you've connected your Discord with YouTube, you have access to the Secret Titans channel. Have an amazing day and I'll see you on Discord.