Imagen 2.5 isnt an improvement
retained text
Image 2.5 has turned out to be a massive letdown in essentially every way. I don’t know how you guys use it, but its “precision editing” essentially does not exist. Leading up to the release of Image 2.5, I was making massive batches of art for an app. It’s anime slop, yes yes. And when I saw 2.5 for the first time I was SUPER excited. Then I started generating with it and... it’s worse. The first image usually looks great. Cool. Then you try to actually make anything consistent with it. It NEVER adheres to minor details, references, an established art style, or even a giant meticulously curated MD telling it exactly what to do. That is by FAR the biggest flaw with image generation right now. I can give it the most curated reference image and prompt known to mankind and eventually I’m still getting the same generic anime girl again. Blush. Loose hair. Those stupid eyes. Random details I never asked for. Different proportions. Different shading. Different everything. It does not matter how specific you are. Image generation is USELESS for batch work and consistency right now. And reiterating on an existing image is somehow even worse. You can have an image that is 95% perfect, ask it to change ONE thing, and it immediately fucks the entire image. This applies to basically every image generator I’ve used, not just OpenAI. They are all obsessed with making a NEW image instead of preserving an EXISTING one. There is a gigantic difference between: “make me a pretty anime girl” and “this is an established character with an established art style. Put THIS character into THIS pose, preserve THESE exact details, and DO NOT TOUCH anything else.” And until OpenAI or somebody else actually locks the fuck in and figures out reference adherence, style locking, character consistency, and ACTUAL precision editing, image generation is basically useless for this entire category of work. Who is the target audience here? Because it sure as fuck isn’t someone who needs to generate 10+ images of the same characters and have them actually look like they belong in the same project. I cannot imagine anyone doing serious batch work with these models and being happy with the result. Right now it still feels like the main use case is: “make me Studio Ghibli!!” 🤓 let me change one character’s shirt without rebuilding the entire fucking image. (or getting flagged 100 times over because covered boobs = fraudulent or scam activity??)
document fetched 2026-09-21 00:20 · canonical https://reddit.com/r/ChatGPT/comments/1wljlam/imagen_25_isnt_an_improvement