By Xiaohongshu · Open Source

Open Source SOTABilingual CN/ENApache 2.030GB VRAM

FireRed Image Edit

A general-purpose image editing model that delivers high-fidelity and consistent editing across a wide range of scenarios. Open-source SOTA with accurate instruction following, high image quality, and consistent visual coherence.

0ImgEdit Score

0BTraining Samples

0sInference Speed

By RedNote · Open Source

FireRed Universal Image Editing

A universal image editing model trained on 1.6 billion samples, achieving state-of-the-art high-fidelity editing across object manipulation, style transfer, virtual try-on, photo restoration and more. Open source under Apache 2.0.

Start Editing

High-Fidelity Editing

Precise instruction following with visual consistency

Lightning Inference

8-step Lightning mode, 4.5s per edit

Bilingual Support

Native English & Chinese instructions

Multi-Scenario

10+ editing capabilities in one model

Key Features

State-of-the-art editing performance with ultimate engineering optimization

State-of-the-Art Identity Consistency

Open-source SOTA in character identity preservation, ensuring subjects remain recognizable across complex edits

Multi-Element Fusion

Freely combine 10+ elements with Agent-powered automatic cropping and stitching — no more struggles with short prompts

Comprehensive Portrait Makeup

Dozens of styles from professional beauty retouching and yellow/olive skin tone brightening to Halloween witch makeup and creative looks

Text Style Reference

Maintains high-fidelity typography and stylized text comparable to closed-source solutions

Professional Photo Restoration

High-quality old photo repair and enhancement with superior detail recovery

Official Showcase

FireRed Image Edit Showcase

Explore the four core editing capabilities of FireRed: portrait editing, multi-image fusion, portrait makeup, and text style reference. All examples are from official documentation.

Portrait Editing

Complex portrait editing including background replacement, clothing changes, pose adjustment, and accessory modification

Official Editing Instructions

Try it now

Open-Source SOTA

Benchmark Results

FireRed Image Edit establishes a new state-of-the-art among open-source models on ImgEdit, GEdit, and REDEdit benchmarks

Open-Source Models6 models

Open-Source Models
Model	ImgEdit_O ↑	GEdit_O ↑ (EN)	GEdit_O ↑ (CN)	REDEdit ↑ (EN)	REDEdit ↑ (CN)
Step1X-Edit-v1.2	3.95	7.480	7.467	—	—
Qwen-Image-Edit-2509	4.31	7.480	7.467	3.99	4.00
FLUX.2 [Dev]	4.35	7.413	7.278	4.07	4.05
LongCat-Image-Edit	4.45	7.748	7.731	4.12	4.12
Qwen-Image-Edit-2511	4.51	7.877	7.819	4.23	4.18
FireRed-Image-Edit	4.56	7.943	7.887	4.26	4.33

FireRed Image Edit official benchmark evaluation chart — Official Evaluation

Data Highlights

FireRed by the Numbers

Trained at scale for production-grade image editing

ImgEdit Overall Score

GEdit Score (EN)

End-to-End Inference

0GB

VRAM Requirement

Testimonials

Community Feedback

What researchers and creators say about FireRed Image Edit

“FireRed's identity consistency in v1.1 is remarkable. Face and character preservation across edits rivals closed-source solutions, and the open-source availability accelerates our research.”

FAQ

Frequently Asked Questions

Common questions about FireRed Image Edit

: A general-purpose image editing model by Xiaohongshu's Intelligent Creation Core Technology Team, built on Diffusion Transformer architecture with Qwen2.5-VL as vision-language encoder.
: 10+ categories: object add/remove/replace, attribute adjustment, background editing, style transfer, text editing, photo restoration, multi-image editing, virtual try-on, portrait makeup, and multi-element fusion.
: 30GB VRAM with optimized inference (distillation + quantization + static compilation), ~4.5s per sample.
: Open-source SOTA on ImgEdit (4.56), GEdit EN (7.943), GEdit CN (7.887), REDEdit EN (4.26), REDEdit CN (4.33), surpassing some proprietary models.
: Yes, native bilingual support for both Chinese and English editing instructions.
: Automatic multi-image processing: ROI detection → crop & stitch → recaption. Supports 1-3 native input images, and 3+ via Agent.
: Yes, full LoRA training code is released. Also provides LoRA Zoo with pre-trained styles (Makeup, Covercraft text style, etc.)
: Apache 2.0, fully open source. Available on HuggingFace, ModelScope, and GitHub.

Try FireRed Now

Experience FireRed Image Edit

Start editing your images with state-of-the-art AI technology

Get Started

10,000+ users

FastFree TrialNo Card

Friend Links

Friend Links · 2