PixelDiT2 edits images pixel by pixel and scores an FID of 1.46 at 256×256
The University of Rochester and NVIDIA built the model to change one detail without disturbing the rest, and it is headed to NeurIPS in December.
In brief
An image AI that edits one pixel at a time
PixelDiT2 works directly on pixels. It reports an FID of 1.46 at 256×256. The aim is fixing one detail without touching the rest. It is research, not a consumer tool.
PixelDiT2 is an image-generation model that works directly on pixels, the tiny dots that make up a picture. It comes from the University of Rochester and NVIDIA, and the team is led by Jiebo Luo. We know this from the university's announcement, the arXiv paper and the project page.
The numbers
The paper reports an FID of 1.46 at 256×256 resolution after 600 epochs of training. At 512×512 it reports 1.48 after 680 epochs. FID is a score for image quality, and lower is better.
The university says the model reaches almost the same performance as the latent-space models behind today's top image generators.
Why it matters
Imagine you ask an AI to fix the glasses on a face, and it also changes the background. Many people have hit this kind of problem with tools like ChatGPT or Google Gemini, as the university's release describes it.
The release blames the way most generators are built. They work in latent space, a compressed version of an image that is fast to handle. Luo says this cuts computation but costs control, because individual pixels are not adjusted.
Background
PixelDiT2 follows an earlier model, PixelDiT. That one was presented at CVPR 2026 and was a best-paper finalist, one of 15 among more than 4,000 accepted papers. The release says it let users target regions, such as the glasses on a face or the color of a car.
What comes next
The team will present the PixelDiT2 paper at NeurIPS this December. The university also says pixel-level control could shape future AI video generation.
Limits
This is a research result, not an editing tool you can download. Sources differ on the framing. The university presents pixel-level control as achieved. The small site superpowerdaily.com notes that FID measures image quality, not whether unrelated details stay untouched. Independent mainstream coverage was not found, so the claims rest mainly on the university, the paper and the project page.