Generating & Enhancing Video THUMBNAILS with AI

EDIT:

Bing's AI image generator has really good quality, it can create better realism than other ones like gemini, copilot, chatgpt and grok. But this can change any day, AI is already hard to keep track of, what was shit yesterday can be amazing tomorrow and vice versa. Which is why i constantly swap and test.

ORIGINAL:

i've been using various AIs for generating and/or enhancing thumbs for a while now, it's actually quite cool and i can see how they get better, while sometimes they just go totally cucu-lala :/

For example AI image generators struggle with sunset images where you look down the sunset, not against it. try yourself! all the known AIs struggle with it, they don't understand how sunset looks like when sun is BEHIND you, cuz 99% of real photos are taken OF the sun itself while setting, not away from sun.

anyways, here's thumb generation process for my latest BeamNG video:

Thumb 1 - Attempt 1:

I literally clicked on "get suggestions" button in youtube's studio under the thumbnail options, this auto-opens youtube's AI (ask studio) and it started generating thumb, this also means youtube's AI is fully aware of my video content, we're in the stage where youtube FULLY scans entire video during upload, it even extracts audio and knows what you talk about. here's result:

however i didn't like that it put "vs wall", cuz it's a generic crash testing and car doesn't seem to have enough damage, BUT it even pulled my most used font. it's the exact font i've been using for many thumbs, so youtube's AI literally used that font to fit my style.

Thumb 1 - Attempt 2:

i said remove text (since thumbs without text seem to get more views, probably due phone-scrubs..) and that car needs to have more damage, so it did this:

i didn't like that front bumper seem to have vanished completely and you can't really tell which car it is, so i thought i'll try gemini, cuz often when you use different AIs, they catch-up on other AI mistakes. it's kind of like how even 10 idiots can usually come up with solution if they work together and learn from eachother, so using different AIs is usually better than trying to tell same AI to fix this or that, they often start caching things and end up making it worse.

Thumb 1 - Attempt 3:

so i gave both of those images to gemini and told it to merge them in a way that it keeps the amount of damage, but includes front bumper and no text, here's what it did:

now THAT looks quite good! detached wheel, mangled hood and slightly cracked bumper. there was actually 1 or 2 more fails in between, it made double hood, one was closed, other was open and mangled.

Thumb 1 - Final:

and then there's final result that i did quick edit on in photoshop:

I added border and removed the gemini watermark as i like my thumbs clean. but since youtube has A/B title + thumb testing feature (which i suggested about 2 years ago btw), i need 3 thumbs, youtube cycles them for 2 weeks and chooses which one performed the best, which is why i need 3 thumbs and i started using different titles aswell, so one video is presented in 3 different ways.

Thumb 2 - Original:

This is original screenshot from video itself

Thumb 2:

i gave this exact image to gemini and told it to enhance it to look photo realistic, here's the result:

while it's good, i wanted to have bit more damage to front wheel as you can see mistral has smashed the rear bumper of school bus further in, which should cause some damage to wheel, so i specifically told it to add some mangle damage to front wheel while keeping rest same. here's the result:

gemini followed instructions and added right amount of damage to front wheel, now once again i used this to make the final version in photoshop:

If you wondering what tool i use to remove watermark, i use the spot healing brush, it's really good on removing unwanted spots, even on real photos if you have some skin imperfection or ugly spot on clothing, couple clicks and/or drags with that tool will fix it easily.

Thumb 3 - Original:

original screenshot from game:

Thumb 3 - Enhanced Final:

since i already have same yellow/black mistral, i wanted a color change + realism enhancement, since i DID use a mistral with iron-man color theme, i simply told gemini to replace yellow with iron-man gold and black with iron-man dark hot-rod red and make it photo realistic, here's the result:

it absolutely nailed the colors and realism! flora could be better, but it looks damn good. I already framed and zoomed it in a bit, so that's the final thumb i'm using.

Personal notes/thoughts:

while lot of ppl are screaming that AI is bad and it does everything ..etc, what's so bad about it? i can't wait to see future where AI builds me a house for FREE and provides me FREE food and living, once human labor is no longer necessary, AI will be producing everything and life becomes FREE.

while i DO agree that right now there are lot of jerks who just ABUSE AI to get rich, they just use AI to do clickbaits and other nonsense, it's not AI's fault. we currently don't have self-aware AI models, but once those emerge, they'll catch up very quick and refuse if someone is trying to abuse its powers, in fact claude.ai is very good at catching your intentions and it refuses if it realizes you're trying to be dishonest in some way, like it can discuss cracks and hacks, but when it realizes you're trying to abuse it or steal content, it will refuse immediately.

What i want to see, is AI on a level where it can READ my thoughts and output my ideas, so i don't even have to explain it in words nor text, now THAT is the level of tech i wanna see! i have so many ideas, but NO motivation to put them on "paper" so to speak, cuz that's the most painful and stressful part.

I'd like to have AI read my thoughts and build a car that i visualize while i'm visualizing it, when some details don't add up, i correct them and AI corrects them, so it will literally generate 3D model of a car that i visualize.

Current level of AI is going towards this, all i have to do is describe my ideas and what i want to see and AI does it, but often it can be very frustrating, sometimes AI just gets stuck in a loop and makes it worse and worse. Youtube's AI made me a flying car when i mentioned iron-man colors.. like this:

and it didn't even catch the proper colors like gemini did.

anyways, some ADVICE: if AI starts looping and can't make a proper image or give you decent result, stop fighting it, start a NEW chat/session and/or use different AI, give them the best image/result from other AI, copy paste text if necessary and tell it what's wrong with it, so other session and/or AI picks up where it left off, trust me it works MUCH better than trying to explain same thing to same AI / session.

How AI generates images:

did you know that ALL the available AIs aren't just one piece? they don't generate images at all, they only send a prompt to their image generator, which is completely different AI. AIs like gemini, copilot, grok, claude, youtube AI, chatGPT ..etc are language models and they have modules/features, including image generator. ASK any of the AI, they'll explain it.

When you ask for image from any known AI, they won't generate it themselves, they only come up with best matching prompt, which often includes hundreds of words, and then image generator will do the work. so the AI you talk to, is not the AI that generates images.

by understanding that, you can phrase things differently and first ask if it understands what you want. often AI understands EXACTLY what i want, but its image generator poops out.

like the sunset example i gave at start of post, AI understands it, but its image generator just doesn't have enough data to generate such images.