It's like 1964 but corporate enforced. Now there are tasks that you are not allowed to do despite being legal.
In the same way, using gpt5 is now very unbearable to me as it almost always starts all responses of a conversation by things like: "Great question", "good observation worthy of an expert", "you totally right", "you are right to ask the question"...
>In the same way, using gpt5 is now very unbearable to me as it almost always starts all responses of a conversation by things like: "Great question"
User preference data is toxic. Doing RLHF on it gives LLM sycophancy brainrot. And by now, all major LLMs have it.
At least it's not 4o levels of bad - hope they learned that fucking lesson.
OpenAI are in a difficult position when it comes to global standards. It's probably easier to see from outside of the United States, because the degree to which the historical puritanism has influenced everything is remarkable. I remember the release of the Watchmen film and being amazed at how pervasive the preoccupation with a penis was in the media coverage.
Imagine if we woke up tomorrow morning and grep refused to process a file because there was "morally objectionable" content in it (objectionable as defined by the authors of grep). We would rightly call that a bug and someone would have a patch ready by noon. Imagine if vi refused to save if you wrote something political. Same thing. Yet, for some reason, we're OK with this behavior from "certain" software?
None of the templates included with e.g. Word were for smut.
Word allowed you to type in smut, but it didn’t produce smut that wasn’t written by the user. For previous enterprise software, that wasn’t really a relevant question.
So… I don’t think it is obvious that the “Word lets you type in smut” implies “ChatGPT should produce smut if you ask it for smut.”
I guess precedent might imply “if you write some smut and ask it to fix the grammar, it shouldn’t refuse on the basis of what you wrote being smut”?
Photoshop, MS word.
But the image editing page linked at the top is more recent, and was added sometime in September. (And was presumably the intended link) I hadn’t read that page yet. Odd there is no dates, at first glance one might think the pages were made at the same time.
SEO guys convinced everyone that articles without dates do better on search engines. I hope both sides of their pillow is hot.
“Image editing” is a curious term, as it appears the site/topic is actually all about generating new images. The term in my mind should be for actual editing of existing, real, images, Eg “remove the coffee table” from this living room photo after uploading the image. I’ve found the actual “image generation” models to be bad at this because they introduce too many artifacts that weren’t in the original, which makes sense because they are really geared for creating images out of thin air.
Multimodal models like qwen3-vl-30b-a3b, however, seem to do quite well with editing existing images without trying to constantly add in new things or trying to change the image in ways that you don’t want, as if it’s trying to do the “lets just generate a new image” thing. imagegpt.com is also good for editing existing images, but not sure what model they are using on the backend.
WRT to Qwen3, is it possible that the API/site you were using was passing your "image edit requests" to something like Qwen-Edit [1] under the covers?
To my knowledge, Qwen3-VL (Vision Language) isn't capable of generating/modifying images - it's purely for doing reasoning about images.
Input: bald man Prompt: give bald man hair Output: edited original, now with hair
That looks like editing to me.
Or are we strictly adhering to the ‘generating new images’ definition because these models technically recreate the entire image? It would be like editing a photo in Photoshop. If you hit “Save” you edited the photo. But if you hit “Save As” and create a new file, the photo wasn’t edited but created as a new image?
The other stuff is text to image (not editing)
Is also pretty obvious that the models have some built in prompt system rules that makes the final output a certain style. They seem very consistent
It also looks like 40 has the temperature turned way down, to ensure max adherence, while midjourney etc seem to have higher temperature.more interesting end results, flourishing, complex Materials and backgrounds
Also what's with 4o's sepia tones. Post editing in the gen workflows?
I don't believe any of these just generate the image though, there's likely several steps in each workflows to present the final images outputted to the user in the absolute best light.
I've done this enough to suspect that most hosted image models don't increase their running costs to try and get better results through additional passes without letting the user know what they are doing.
Many of the LLM-driven models do implement a form of prompt rewriting though (since effectively prompting image models is really hard) - some notes on how DALL-E 3 did that here: https://simonwillison.net/2023/Oct/26/add-a-walrus/
https://en.wikipedia.org/wiki/Space_hopper#/media/File:Space...
The title of this article is "image editing showdown", but the subject is actually prompt adherence in image generation from prompting.
Midjourney and Flux Dev aren't image editing models. (Midjourney is an aesthetically pleasing image generation model with low prompt adherence.)
Image editing is a task distinct from image generation. Image editing models include Nano Banana (Gemini Flash), Flux Kontext, and a handful of others. gpt-image-1 sort of counts, though it changes the global image pixels such that it isn't 1:1 with the input.
I expect that as image editing models get better and more "instructive", classical tools like Photoshop and modern hacks like ComfyUI will both fall away to a thin fascade over the models themselves. Adobe needs to figure out their future, because Photoshop's days are numbered.
Edit: Dang, can you please fix this? Someone else posted the actual link, and it's far more interesting than the linked article:
https://genai-showdown.specr.net/image-editing
This article is great.
https://generative-ai.review/2025/09/september-2025-image-ge...
UPD: suggested here https://github.com/scpedicini/genai-showdown-public/discussi...
I don't fully understand the iterative methodology tho - they allow multiple attempts, which are judged by another multimodal llm? Won't they have limited accuracy in itself?
What is the metric on which these models are being judged?
It's hard to define a discrete rubric for grading at an inherently qualitative level. To keep things simple, this test is purely PASS/FAIL - unsuccessful means that the model NEVER managed to generate an image adhering to the prompt. For example, Midjourney 7 did not manage to generate the correct vertical stack of translucent cubes ordered by color in 64 generation attempts. In many cases, we often attempt a generous intepretation of the prompt - if it gets close enough, we might consider it a pass.
Put another way: if I were to show the final image to a random stranger on the street, would they be able to guess what the original prompt was? (aka the Pictionary test).
To paraphrase former Supreme Court Justice Potter Stewart, "I may not be able to define a passing image, but I know it when I see it."
Now, are LLM judges flawed? Obviously. But they are more shelf stable than humans, so it's easier to compare different results. And as long as you use an LLM judge as a performance thermometer and not a direct optimization target, you aren't going to be facing too many issues from that.
If you are using an LLM judge as a direct optimization target though? You'll see some funny things happen. Like GPT-5 prose. Which isn't even the weirdest it gets.
If this one were shown in a US work environment, I might say a collegial something privately to the person, about it not seeming the most work-appropriate.
The ability to clearly understand the image being edited, and make edits that look natural to that understanding, are far beyond any of the other models.
I am biased on this since I built it and it officially launches on Friday 10/31
I'm pretty sure that only Gemini made it. Other models did not meet the 'each tentacle covered' criteria.
Did current models overcome the 10:10 bias?
Three tuple: (original image, text edit instruction, final image).
Easy to patch for editing models, anyway. Maybe not text to image models.
It probably comes up more than you think. Storyboarding, product placement, model images, etc.
It's not critical in the short term, but it'll wind up on their backlog for sure.
A prompt id love to see: person riding in a kangaroo pouch.
Most of the pure diffusion models haven’t been able to do it in my experience.
Edit: another commenter pointed out the analog clock test, lets add the “analog clock showing 3:15” as well (: