@willfrancis24: Is ChatGPT Images 2.5 any good? Or just another AI slop image generator? Let's take a look at whether it can overcome the typical problems of previous generations of AI image tools. I'll show you what features I think are most useful. So this came out on Tuesday. It's on the free version too, and OpenAI say it's sharper, faster and better at editing. Of course, the first test is a poster for the village fair. Just imagine if it did this really well, because we've all seen too many of these AI slop posters. What happened to good old Comic Sans in a Word document? This was with the previous version of ChatGPT. And here's the new version. I do think this will still make people angry. It's nicer. It's more subtle. The font is very recognisably AI. To be fair, in the new templates here is a poster template, and here are the results you get from that, with a reference image of one of those cool posters for an art exhibition. When we use that, it looks really cool. Next test is that weird yellow tint they add to make things look more organic, it’s become a real AI tell. In fact, someone built unyellowgpt.com just to remove that grading from AI images. And yes, we can see that it has stopped doing that, which is great. The third test is detailed scenes. Here's a Where's Wally (Waldo) puzzle, and yes, Wally is in here once. I've had a good look around it, and everything seems good, which is quite impressive. The third test is detailed scenes. Here's a complex, asymmetric view of a city with life on every corner and things happening on every block. It's 20 by 20 blocks in size. Wow, that is seriously detailed. The fourth test is complex text combined with real-world knowledge. This is a tough test, but also a common use case. If we want it to make infographics, posters, etc., we want that information to be correct. And it wasn't great at that, unlike Gemini. This infographic of London's tallest buildings is a lot better now, and the numbers are all correct. The last test is: can it break away from the conventions it's trained on? Can it show a clock that doesn't read 10:10? Can it show a wine glass that isn't perfectly half full? It used to struggle with that, however much you asked it, but it seems to have overcome that. That's great because it means that it's less influenced by the averageness of its training data and more influenced by what you want. Great. For me, the most useful upgrades are: Designs with text, like the header for my weekly newsletter, which you can subscribe to, see my link in bio. I've noticed it's much better when it uses my actual fonts rather than its own telltale AI typography. Consistency of character or object. If you give it a photo of a person, maybe you, and ask it to put you in 20 different outfits, it can do that, but it will still look exactly like you. You might put a product that you sell into 5 different settings, and it's great at not changing the product. There are very functional uses for it. This sketch feature: you get a little canvas to draw in. This isn't for bringing doodles to life, although your kids might like that. This is for influencing the layout of what it produces, which is especially good for design tasks. It is very interesting that OpenAI have just blocked design companies like Adobe from advertising in ChatGPT. I think they are competing with those kinds of companies now with this product. #chatgpt #openai #chatgpttutorial #ai #llm #Tech #technology #learnontiktok #aitools #artificialintelligence

Will Francis - AI + Marketing

Open In TikTok:

Region: GB

Friday 11 September 2026 20:45:34 GMT