09:23Dill Toma · Tutorial
The Thumbnail Secrets Nobody Teaches You
A ten-year thumbnail designer breaks his process into four pillars: audience, attention, clarity, and desire.
September 19thThe argument in one line.
A three-tool loop, ChatGPT for prompts, Freepik for AI image generation, and Photoshop for compositing, lets a thumbnail designer build a photorealistic custom thumbnail without ever using a real photo of the subject in the final image.
Read if. Skip if.
READ IF YOU ARE…
- You design thumbnails for your own channel or for clients and want to speed up the concept-to-final pipeline without hiring an artist.
- You're already comfortable in Photoshop and want to add AI image generation as one more tool in that process, not replace it entirely.
- You want a concrete, step-by-step look at prompting ChatGPT and Freepik together instead of a general 'AI can do thumbnails now' overview.
SKIP IF…
- You're looking for a no-editing-skill shortcut. Every AI output here still goes through manual Photoshop compositing, masking, and color grading.
- You want AI tool comparisons or pricing breakdowns. The video shows one specific workflow, not a buyer's guide.
The full version, fast.
Dill Toma rebuilds the exact thumbnail used on this video, live, using only three tools. ChatGPT writes and revises the image prompt from reference photos. Freepik generates the raw images, he tests multiple AI models and feeds screenshots of bad results back into ChatGPT to sharpen the next prompt. Photoshop does the real work: Select Subject and Generative Fill remove unwanted elements, pieces from different AI generations get composited into one base image, a rotated head gets reconnected to the neck with Generative Fill, shadows get hand-painted back in for realism, and a final color grade unifies everything into one coherent photo.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:37
01 · Hook + credibility
Years of thumbnail design experience stated, then the claim AI changed the whole process. Promises to build the exact thumbnail on this video live.
00:37 – 01:29
02 · The three-tool stack + starting in ChatGPT
Names the only three tools used (ChatGPT, Freepik, Photoshop), then uploads reference photos into ChatGPT to start building the image prompt.
01:29 – 02:39
03 · First Freepik generation with Nano Banana Pro
Picks an AI image model inside Freepik, uploads the same reference photos in order, pastes the ChatGPT prompt, and generates the first batch of images.
02:39 – 04:11
04 · Prompt revision loop + the screenshot trick
Several rounds of going back to ChatGPT to revise the prompt. Introduces feeding Freepik screenshots back into ChatGPT so it can critique its own failed output, plus a mid-video pitch for 1-on-1 consulting.
04:11 – 05:09
05 · Photoshop base image: Generative Fill removes the subject
Picks a base AI image, uses Select Subject plus Generative Fill to erase the person and laptop from it, clearing a plate to build the composite on.
05:09 – 06:15
06 · Compositing the hand + swapping the laptop angle
Masks in a hand from one AI image and repositions it, then sources a better laptop angle from a different Freepik generation.
06:15 – 07:25
07 · Positioning the laptop + painting shadows back in
Places the new laptop asset, then hand-paints contact shadows from the body and desk back onto the mask for realism.
07:25 – 08:48
08 · Fixing the head angle with Generative Fill
Cuts out the head, rotates it to a straighter angle, masks the old head, then uses Generative Fill to reconnect the neck without distorting the face.
08:48 – 09:44
09 · Expanding the background + nailing composition
Uses Generative Fill to extend the background up and to the left, then adds guidelines to check the composition is centered.
09:44 – 10:15
10 · Color grading pass
Adds a Levels and Hue/Saturation adjustment on a shirt mask to darken and desaturate it, plus a Color Balance layer over the whole image.
10:15 – 11:06
11 · Faking a phone notification via ChatGPT
Uploads the near-final composite into ChatGPT and asks it to add a text-box notification graphic above the laptop without changing anything else, then skews and positions the result in Photoshop.
11:06 – 11:58
12 · Final composite + sign-off
Adds a green glow, final touches, reveals the finished thumbnail, and closes with a subscribe ask.
Lines worth screenshotting.
- Feeding a screenshot of a failed AI image generation back into ChatGPT and asking what's wrong produces a sharper prompt revision than describing the problem in words.
- Uploading 2-3 reference photos into ChatGPT before asking for a prompt keeps the AI-generated face consistent with the real subject across multiple Freepik generations.
- A single AI-generated image is rarely the final asset. The real output of each Freepik generation is raw material for a Photoshop composite, not a finished thumbnail.
- Photoshop's Select Subject plus Generative Fill can cleanly erase a full person and the objects around them from a photorealistic AI background in one pass.
- Building one thumbnail from pieces of several different AI generations, a hand from one image, a laptop angle from another, gives more control than regenerating the whole scene.
- A rotated head can be reconnected to its original neck using Generative Fill on the leftover area, which blends the geometry without redrawing the face.
- Hand-painting contact shadows back onto a mask is what stops a technically clean composite from reading as pasted-in.
- Generative Fill can expand a canvas outward, which is a fast way to recenter a composition after cropping and compositing have thrown it off balance.
- Darkening or desaturating one competing element in frame, like a shirt, can reduce visual clash with a bright object nearby more effectively than grading the whole image.
- ChatGPT can add a convincing fake UI element, like a phone notification, directly onto an uploaded photo when told explicitly not to change anything else about the image.
A repeatable three-tool loop for AI thumbnails
Thumbnail workflow
The real skill is chaining ChatGPT's prompt-writing, Freepik's generation, and Photoshop's compositing into one loop where each tool fixes what the last one got wrong.
02The three-tool stack + starting in ChatGPT
- The entire workflow runs on exactly three tools: ChatGPT for prompt-writing and visual critique, Freepik for AI image generation, and Photoshop for compositing.
- Uploading 2-3 reference photos into ChatGPT before asking for a prompt keeps the AI-generated face consistent with the real subject across multiple Freepik generations.
03First Freepik generation with Nano Banana Pro
- Freepik lets you switch between AI models and upload reference images in the same order they went into ChatGPT, which keeps the face consistent across the batch.
- Treat a first batch of AI generations as a direction check, not a finished asset.
04Prompt revision loop + the screenshot trick
- When an AI output isn't working, revise the prompt in ChatGPT rather than regenerating blind in Freepik. The prompt is the lever, not the seed.
- The underused trick: screenshot the bad AI generations and feed the images back into ChatGPT, asking what's wrong. A visual reference produces a sharper revision than a text description.
- You can chain generations by feeding a prior output back in as a reference image for the next prompt, instead of starting over each time.
05Photoshop base image: Generative Fill removes the subject
- Pick one AI generation as the base image, then use Select Subject plus Generative Fill to cleanly erase what you don't need rather than regenerating the whole scene.
- Generative Fill can remove a full subject and the objects around them from a photorealistic background in one pass, leaving a clean plate to build on.
06Compositing the hand + swapping the laptop angle
- Build the final composite from pieces of multiple AI generations, not one single output. A hand from one image, a laptop angle from another, masked together.
- Because the source images came from the same reference photos and prompt chain, lighting and skin tone stay consistent enough that cut pieces drop in with minimal blending work.
07Positioning the laptop + painting shadows back in
- Grab a second AI generation specifically to source an asset the first one got wrong, instead of trying to fix the flaw in Photoshop by hand.
- Hand-painting contact shadows onto a new mask is what stops a technically clean composite from reading as pasted-in.
08Fixing the head angle with Generative Fill
- To fix a subject's head angle after compositing, cut the head out, rotate it, then run Generative Fill on the leftover neck area to reconnect it without distorting the face.
- Spot-check that an AI fill preserved identity, hairline and face shape, before moving on. Not every generative edit holds proportions correctly.
09Expanding the background + nailing composition
- Generative Fill also expands canvas, not just removes objects. Pulling the background wider re-centers a composition after cropping and compositing have shifted it off balance.
- Add visible guidelines before judging composition by eye. What looks centered without guides often isn't.
10Color grading pass
- A final grading pass, Levels and Hue/Saturation on a masked layer plus a Color Balance adjustment over the whole image, is what makes mismatched source images read as one coherent photo.
- Darkening or desaturating one competing element in frame can reduce visual clash with the brightest object more effectively than adjusting the whole image.
11Faking a phone notification via ChatGPT
- ChatGPT can add a convincing fake UI element onto an uploaded image when told explicitly not to change anything else about the photo, which stops it from redrawing the rest of the scene.
- Treat an AI-added UI element as a draft. Pull it into Photoshop to resize, skew to match the surface angle, and add a glow so it reads as flush rather than floating.
Terms worth knowing.
- Generative Fill
- A Photoshop AI tool that fills a selected area of an image by generating new photorealistic content matching the surrounding scene, used here to remove subjects and to reconnect a rotated head to its neck.
- Select Subject
- A Photoshop AI selection tool that automatically outlines the main person or object in a photo, used as the first step before masking or Generative Fill.
- Freepik
- A design asset platform whose image generator lets users choose between multiple third-party AI image models and upload reference photos to guide the output.
- Nano Banana Pro
- An AI image generation model made by Google, available inside Freepik, used in this video for realistic rendering of faces from reference photos.
Things they pointed at.
01:35toolNano Banana Pro (Google, via Freepik)
Lines you could clip.
00:08
“AI has completely changed how I design thumbnails.”
03:00
“This is a pro tip that I don't see a lot of people doing: taking screenshots of the generations by Freepik and throwing it back into ChatGPT and asking what's wrong with these images.”
11:20
“It's honestly insane how all of this was built just from a few AI tools, and the crazy thing is I didn't even have to use an actual picture of my face.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
I've been designing YouTube thumbnails for years and I've worked with some of the biggest creators on the platform and I can confidently say ai has completely changed how i design thumbnails in this video i'm going to show you the exact workflow i use to create high performing thumbnails faster and easier than ever before i'm going to walk you through all the tools i use on real client projects and i've never broken it down as simple as this by the end of this video you'll be able to copy my exact process to create better thumbnails faster than ever and by the way the thumbnail you clicked on in this video is the one we're going to be creating together live so if you want more thumbnail education videos like this then subscribe now we are only going to be using three tools chat gbt free pick and photoshop let's first start with chat gbt and build the concept for this thumbnail once inside chat gbt go ahead and add in reference images of yourself or your client
Simply come to the plus, click add photos and I'm going to add in three photos that are found on my YouTube channel. Since we're designing a thumbnail for this video, let's go ahead and give chat some details about our video. So first I'm going to start off with I need a prompt for free pick.
Now once we have this prompt written out to ChatGPT, let's click enter and see what it gives us. And after just a few seconds, ChatGPT gave us a full prompt that we can use inside of free pick. So let's jump over to free pick and generate some images.
once inside freepik we can head over to the image generator now this is the cool thing about freepik you can choose from a ton of different ai image models for this example we're going to be using google nano banana pro and as you can see the results are so realistic it's insane make sure we're on 16 by 9 and we have ai prompt enabled now let's head over to upload and upload our reference images in the same order that we uploaded them into chat gpt and now as you can see here we have image one image two and image three let's go ahead and paste in our prompt and generate all right so these were the results that freepik gave me And at a first glance, they're not bad, but they're just not exactly what I want.
So let's go back into chat GPT and revise our prompt a little bit. All right, after some simple revisions and a couple more seconds, chat GPT is giving us another full prompt that we can throw into free pick. And let's go ahead and paste it in and generate four more images.
Okay, I like these first two the most. This is a cool perspective, but I don't think it's gonna work for this. This just feels...
A little too bland. This feels a lot more dynamic. so let's go back to chat gbt and revise the prompt just a little bit more okay i gotta be honest i'm not a fan let's take a screenshot of the images that were generated let's uh scribble these ones out and then let's throw this into chat gbt now this is a pro tip that i don't see a lot of people doing is taking screenshots of the generations by free pick and throwing it back into chat gbt and asking what's wrong with these images or giving chat gbt a visual to see what you're actually talking about it seems to help a lot so once again let's copy this prompt and paste it into free pick all right once again not a fan of these images i'm actually going to take this thumbnail and feed it back into chat gpt and see what we can do with that so now i'm going to say use this image as a image one now we can go back into free pick and upload just this image paste in our new prompt and let's see what results it gives us this time okay i'm gonna go back to these two thumbnails i'm gonna download them both and now we're gonna go into photoshop quick pause if you want me to personally help you with your thumbnails i offer one -on -one consulting calls where we go through your channel and fix your thumbnails together the link to book a call or work with me directly is in the description also make sure you're following me on my other socials for more tips and breakdowns i don't post here
back to the video okay so this image is going to be our base for the thumbnail but i want to take some assets from this thumbnail and just merge them together so the first thing i'm going to do is just drag everything down a little bit let's go ahead and select subject on our base now i can simply use photoshop's ai generative fill to remove me and remove the laptop since we just cut myself out once the area is selected we can click generative fill and just leave it blink and hit generate and as you can see it perfectly removed every asset from the image now let's go to this image and do the exact same thing select the subject and it should select me as well as the laptop perfect and all i want from this image is honestly just my hand here in the foreground So I'm going to remove everything else using a mask.
And there's just my arm. Now if we turn this back on and reposition the arm to be about there, that almost already looks perfect. and just like that the arm fits in perfectly because we're using the same prompts so all the images are going to be similar now i believe the laptop is facing a little bit too much away from the camera so let's go back to free pick and see if we can find a picture that has the laptop facing more towards the camera now i like this image here because of how colorful all the apps on the bottom are compared to this one so let's download this image and cut out just the laptop there we go now we can position this right about where we want it let's go ahead and remove this one And that simply, you can already see such a big difference.
Now let's go ahead and add in some of this shadow back. I'm just painting on the mask and adding some of these shadows back into the image just so that it stays realistic. And once again, I'm going to paint back in the shadows on the surface of the desk that are made by my body.
And there we go. The shadows make it look so much more realistic. Now, what I want to do is I want to actually tilt my head a little more straight up.
So I just simply cut out my head and rotated it slightly to the angle that I want. Now I'm going to mask out my other head just like that. And now I'm going to select the area left over and the part connecting my neck.
And we're going to click generate and see what this does for us. Hopefully it removes the head in the background, but also makes my neck connect to the head that I added in. I mean, that looks pretty spot on.
As you can see, it didn't change the shape of my face too much. Kept the hair. But just made the neck actually connect.
Beautiful. Now, of course, let's not forget to expand the background up here. Perfect.
I'm also going to expand this out a little bit to the left just so that everything can be a little more centered and a little bigger. Let's get some guidelines to make sure I'm doing this correctly. And that looks...
Pretty good composition wise to me. Now, the next thing I want to do is actually add in one of my thumbnails onto the MacBook here. All right.
Now I added one of my thumbnails to the MacBook screen. Let's go ahead and add some finishing touches. All right.
Let's create a levels.
Let's make this the mask. And let's darken my shirt a bit. Let's also create a hue saturation with the same mask and turn down the saturation.
And I feel like the darker shirt just doesn't clash as much with the MacBook, which I like. Let's also add a color balance layer to the whole image and just slightly adjust the colors to our liking. Okay, let's export this image and go back to ChatGPT.
Let's upload our image and say use this image as at image one. Don't change anything about the image. Just add a text box above the laptop that says image generated.
And let's just say give me. prompt and see what we can do with the prompt that it gives us. Again, don't forget to upload your reference image and let's generate.
Okay, I already like how these two are looking. Wow. Okay, this one is really good.
I'm going to save this one. Let's open it in Photoshop. What we're going to do is just cut out the notification just like that.
Let's position it so that it's bigger. And let's also skew it so it looks like it is flush with the laptop. And just like that, this already looks so good.
Let's go over this with a little green glow and see how that looks. It's honestly insane how all of this was built just from a few AI tools. And the crazy thing is I didn't even have to use an actual picture of my face.
Let's go ahead and add the final touches to this thumbnail. This is ready to be exported and uploaded. And that's it, that's exactly how I make thumbnails with AI in 2026.
If you've learned something, subscribe for more thumbnail videos like this and check out my links below to contact me. Thank you for watching, I'll see you in the next video.
The Hook
The bait, then the rug-pull.
Dill Toma opens with the credibility pitch: years of designing thumbnails for some of YouTube's biggest creators, and a claim that AI has completely changed the process. Then he does what most 'AI changed my workflow' videos skip. He builds the actual thumbnail you clicked on to watch this, live, using nothing but ChatGPT, Freepik, and Photoshop.
Named ideas worth stealing.
03:00model
The Prompt, Generate, Critique Loop
- Write the prompt in ChatGPT from reference photos
- Generate images in Freepik
- Screenshot the results that miss
- Feed the screenshots back to ChatGPT and ask what's wrong
- Revise the prompt and regenerate
Instead of regenerating blindly when an AI image misses the mark, the loop treats ChatGPT as the critic: show it the actual failed output and let it rewrite the prompt with that visual context.
Steal forany AI image-generation workflow, not just thumbnails
00:50list
Three-Tool Thumbnail Stack
- ChatGPT (prompt writing + visual critique)
- Freepik (AI image generation, multi-model)
- Photoshop (compositing, retouching, color grade)
The entire video runs on exactly three tools, each doing one job. ChatGPT never generates images and Freepik never does the final composite.
Steal forscoping any AI-assisted design pipeline down to the minimum tool count
How they asked for the click.
VERBAL ASK
03:43product
“Quick pause if you want me to personally help you with your thumbnails, I offer one-on-one consulting calls where we go through your channel and fix your thumbnails together. The link to book a call or work with me directly is in the description.”
Mid-video pattern-interrupt pitch, dropped right after the prompt-revision demo and before switching into Photoshop, framed as a pause rather than a hard sell.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
