Ok guys this is crazy FLUX 3 image precision edit is real! GPT 2.5 is dead in the water

Oct 2, 2026 · 3:41 PM UTC

33
38
772
70,949
To all who asked how drift was calculated with Opus 5.5 Drift = the average change in pixel colour, versus the original photo, in areas no edit ever asked to touch (two wall corners and a strip of floor). Each output is first aligned to the original so framing differences don't count. 0 = untouched, 255 = completely different. Every edit uses the previous edit's output, so the score shows damage piling up.
8
1,643
Sort replies: Relevant Recent Liked
Replying to @mightyking
I also ran into the same issue when repeatedly editing the same image with GPT Image 2.5. In commercial ad design, multiple edits on a single image are pretty common. This made me pay closer attention to the Flux 3.0 model🧐
1
1
25
And you could still apply multi object item edit attempt directly to the original. So my experiment exaggerated the result to prove the point. Look also at the drift after 1 iteration on GPT. Makes the image useless
16
Replying to @mightyking
No ones gonnna pay extra for img edits especially when all anyone wants is porn
1
1
76
Weird argument
1
60
Replying to @mightyking
1
2
1,043
Correct
1
896
Replying to @mightyking
They could never catch up GPT Image, so they have no choice but to pour their efforts into mastering a few specific editing techniques.
1
1
355
Smart approach
299
Replying to @mightyking
Who edits images like this? It doesn’t make sense to edit from a previously edited image. Instead, use the original image and edit it once for all.
2
8
1,480
Even in that case freeze the video and see the insane drift from just 1 edit between the two models
2
1
1,340
more replies
Replying to @mightyking
Why does this happen?
1
1
544
Ask @sama
1
497
more replies
Replying to @mightyking
But isn't Flux 3 for generating videos?
1
1
902
Flux 3 Image released yesterday with new percision editing by coordinates
2
854
Replying to @mightyking
DO NOT MAKE HYPE PLEASE
1
1
138
Where did you see hype?
116
Replying to @mightyking
Is it open source?
1
77
They said they will. Now we wait
1
46
Replying to @mightyking
Will it open source /weight?
1
1
66
Let’s see
20
Replying to @mightyking
the drift is real where does Nana Banana 2, Pro come in?
1
1
273
It doesn’t
235
Replying to @mightyking
GPT has always been garbage. That coffin was sealed when Nanobanana pro came out.
1
2
231
I mean it had good looking people, just annoying noise
167
Replying to @mightyking
I wonder if the degradation of gpt2.5 is because it doesn’t randomize the seed with every edit. Drift gets dramatically worse if you keep the same seed (seen on local models like qwen and Klein edits )
1
1
443
Or they don't focus seriously on the generative part as much as BFL do
370
Replying to @mightyking
GPT
1
99
You like it?
80
Replying to @mightyking
It’s the watermark they keep injecting
1
4
384
you think so?
353
Replying to @mightyking
chatgpt image is just degrading because it is compounding the watermark noise. start from first input image each time, you'll notice that they are on par, chatgpt will get things right more often
1
4
538
Nope - check the level of drift in var 0 to 1 between gpt and bfl
1
457
Load more