FireRed Image Edit 1.1 image model

FireRed-Image-Edit-1.1 is an open source AI model for precise image editing using text instructions which was dropped at the beginning of March 2026. It changes existing pictures instead of making new ones from zero. The system runs on a diffusion–transformer setup and tries to keep the image looking steady. Text stays readable, faces stay the same. And edits look natural.

The model comes from the FireRed AI group inside Xiaohongshu, a Chinese social platform also called RedNote. FireRed-Image-Edit sits inside a bigger group of vision models made by the same team. The focus here is simple. Make AI image editing more steady… you know, the kind of tasks older models often struggled with. Things like changing text in a picture or keeping a person recognizable after style changes.

The system uses a diffusion–transformer design tuned for instruction-based edits. Training used a huge dataset… about 1.6 billion image and text pairs. These included normal text-to-image samples and editing examples. After cleaning the data the team kept over 100 million higher quality samples to train the final model.

A big goal was clear instruction following. The training happened in stages – pre training, supervised fine tuning and reinforcement learning. The developers also added custom tuning methods. These help the model keep text readable and keep identities stable during edits.

FireRed-Image-Edit works as a general editing tool. It can change backgrounds, shift styles, repair old photos or tweak objects while the main scene stays the same. The creators say tests show results that compete with both open source and closed editing systems.

Image editing and generation. The model supports image-to-image edits. It can repair images, clean noise and sharpen details. It also adds color to black and white photos.

Editing tools. It removes objects or unwanted effects. It fills missing areas with inpainting. It can also replace parts of an image with generative fill or swap the background.

Style and identity tools. The model transfers styles like anime or oil painting. It can follow reference images and keep characters looking consistent.

Image improvements. It restores faces, removes noise and artifacts and increases image resolution with upscaling.

Outputs are edited images. Resolution depends on the tool or pipeline used… most results look similar to modern diffusion outputs around 1024×1024.

There are a few extra abilities too. The system can edit text inside images while keeping the font style, angle and lighting. It also supports identity-safe style transfer so people stay recognizable after visual changes. Old photos can be repaired with detail recovery and scratch cleanup.

Running FireRed Image Edit locally needs a CUDA-ready NVIDIA GPU. That’s the main requirement. The system runs best with stronger GPUs that have more memory.

A setup like an RTX 3090 or a newer GPU with around 24GB VRAM works well for smooth runs. Lower GPUs might still run it… but things can get slower or need tweaks.

Typical uses are pretty simple. Edit objects in photos with text prompts. Change the background or environment. Turn pictures into art styles like anime or oil painting. Restore damaged photos. Or change text in posters while keeping the same typography.