FireRed Image Edit is a general-purpose text-guided image editing model open-sourced by the Xiaohongshu Super Intelligence team (FireRedTeam) on February 14, 2026. Based on the Diffusion Transformer architecture, it excels in high fidelity, strong instruction following, and global consistency, achieving SOTA (State-of-the-Art) results in multiple open-source image editing benchmarks, with some metrics approaching closed-source commercial models.

FireRed Image Edit is a versatile image editing model capable of delivering high-fidelity and consistent image editing effects across various editing scenarios, leading in performance among open-source image editing models.

FireRed Image Edit Technical Background: From Global Generation to Local Inference

Traditional image editing models often cause semantic drift in non-target areas (e.g., modifying clothes but changing the background) when performing "local modifications."

FireRed Image Edit introduces an architecture called FireEdit, whose core is the Region Aware mechanism. By combining visual language models (VLM) with diffusion models, the system attempts to better understand user instructions while more accurately locking onto spatial coordinates that need editing. This design aims to balance local modifications with global harmony for complex instructions.

Core Features

Outstanding Editing Performance: Precise instruction following, high-quality image generation, strong visual coherence, leading performance among open-source models

Native Editing Capability: Directly built on text-to-image foundation models, natively endowed with image editing capabilities

High-Fidelity Text Style: Capable of highly restoring text styles within images, with effects comparable to closed-source solutions

Old Photo Restoration: Achieves high-quality restoration and enhancement of old photos

Flexible Multi-Image Editing: Supports flexible editing scenarios like virtual try-on with multiple images

Text Editing

FireRed Image Edit innovatively proposed Layout Aware OCR based Reward.

During the reinforcement learning phase, the team not only penalized typos but also penalized character misalignment, size anomalies, and layout breakdowns.

This enables the model to more accurately edit and maintain the original text style during poster modifications and text replacements.

Example: Modify the text "programme" in the bottom right corner of a poster to "program ongoing," while keeping font and style consistent.

Creativity and Multi-Image Generation

Leveraging the powerful model architecture, FireRed Image Edit supports creative scene generation and multi-reference image generation, enabling style transfer or multi-image fusion.

Benchmark Testing

To verify the model's capabilities, the R&D team constructed the REDEdit Bench benchmark dataset. This dataset focuses on diverse editing scenarios and instructions that better align with human language habits, comprehensively assessing the performance of image editing models. It collected over 3,000 images from the internet, which were screened and designed by experts, ultimately forming 15 categories and 1,673 bilingual (Chinese-English) image editing sample pairs.