🤗 HuggingFace | 🤖 ModelScope | 🖥️ Demo | 📄 Technical Report

FireRed-Image-Edit-1.1 Based on the FireRed-Image-Edit-1.0 foundation model, it optimizes portrait consistency, multi-element fusion, stylized text reference, and portrait makeup effects.

FireRed-Image-Edit-1.0 is a general-purpose image editing model that delivers high-fidelity and consistent editing across a wide range of scenarios. FireRed-Image-Edit delivers leading open-source results with accurate instruction following, high image quality, and consistent visual coherence.

• 🆔 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.

• 🔧 Open LoRA Training Ecosystem: Full training code released for custom style creation, optimized samplers maximize GPU efficiency for identical tasks, sizes, and input counts.
• ⚡ Extreme Speed Optimization: Complete acceleration suite featuring distillation, quantization, and static compilation—delivering 4.5s end-to-end generation with just 30GB VRAM
• 🤖 Intelligent Agent Workflow: Automatic multi-image processing handles complex compositions like virtual try-on without requiring lengthy prompt engineering
• 🔌 Universal Deployment: Native ComfyUI node support and GGUF lightweight format compatibility for seamless production integration
• 🏋️ Efficient Training Methodology: Offline feature pre-extraction completely decouples VLM inference from training workflow, eliminating generation overhead for maximum convergence speed.

• 🏗️ Backbone-Agnostic Architecture: Editing capabilities injected through full Pretrain → SFT → RL pipeline, transferable to any T2I foundation model.

• 2026.03.03: We release FireRed-Image-Edit-1.1, which, based on the FireRed-Image-Edit-1.0 foundation model, optimizes portrait consistency, multi-element fusion, stylized text reference, and portrait makeup effects.
• 2026.03.01: We offer a lightweight inference script (including distilled Lora, quantization, db_cache, and static compilation), now requiring only 30GB VRAM and ~4.5s/sample. 🚀 Try it by simply running python inference.py --optimized True.
• 2026.02.28: We released the Train, supporting HSDP/FSDP, Disaggregated Training, and Multi-Condition Aware Bucket Sampler.
• 2026.02.27: We released the Agent module for instruction rewriting, multi-image preprocessing, supporting automatic ROI detection, image stitching for editing with more than 3 images.
• 2026.02.27: We provided FireRed-Image-Edit-1.0-ComfyUI workflow. Check more details on Huggingface
• 2026.02.14: We released FireRed-Image-Edit-1.0 model weights. Check more details on Huggingface and ModelScope.
• 2026.02.10: We released the Technical Report of FireRed-Image-Edit-1.0.

FireRed-Image-Edit establishes a new state-of-the-art among open-source models on Imgedit, Gedit, and RedEdit, while surpassing our closed-source competitors in specific dimensions—a distinction further corroborated by human evaluations highlighting its superior prompt following and visual consistency. Detailed information can be found here.

Some real outputs produced by FireRed-Image-Edit across general editing.

case1: 将图2的男人，穿着图2的黑色"New York Bears"棒球夹克和迷彩裤子和蓝黑配色的AJ1高帮球鞋，在图1的空旷的橄榄球场上。球场阳光明媚，他带着图2黑色的帽子，帽檐是红色，夹克上的白色条纹在暗光中格外醒目。他的左肩随意地挎着图3那只复古棕色皮质旅行包，包口微微敞开，露出里面图3那个磨损的棕色橄榄球。右手则轻松地拖着图3的白色滑板，板底那威武的黑色狮子图腾在草地映衬下如同猛兽。图3那只壮实的深棕色斗牛犬安静地蹲坐在他脚边的阴影里，忠诚地望向主人。男人脖子上挂着图2那副黑色Beats耳机，地上放着图3那座复古的青铜奖杯，在阳光下下泛着胜利的光芒。整个场景融合了街头潮流与竞技体育的质感，空旷球场、皮革装备的温润光泽、以及滑板带来的街头气息，共同构成了一个关于青春、热爱与赛后孤独的静谧时刻。

case1: 为人物添加欧美Y2K妆：使用冷白皮哑光粉底均匀肤色，描绘粗平的深棕色挑眉，眼部涂抹亮片银灰眼影并晕染至眉骨，画上黑色上扬眼线，粘贴浓密假睫毛，用浅金色高光提亮卧蚕，在苹果肌扫上蜜桃色腮红，唇部涂抹镜面玻璃唇釉，并在颧骨处轻扫修容粉。

case2: 为人物添加缎光底妆：使用自然色缎光粉底均匀肤色，描绘自然眉形并填充浅棕色眉粉，眼部涂抹深棕色眼影并晕染眼尾，画自然内眼线，刷上浓密睫毛膏，在卧蚕处提亮，涂抹水润感红色豆沙色口红并勾勒唇形，在苹果肌扫上粉色腮红，在鼻梁、颧骨处轻扫银色高光。

1. Install dependencies

pip install -r requirements.txt

2. Use the following code snippets to generate or edit images.

python inference.py \
    --input_image ./examples/edit_example.png \
    --prompt "在书本封面Python的下方，添加一行英文文字2nd Edition" \
    --output_image output_edit.png \
    --seed 43

The Agent module provides Recaption & Multi-Image Preprocessing capabilities.

FireRed-Image-Edit natively supports 1–3 input images. When users need to edit with more than 3 images, the built-in Agent module automatically:

1. ROI Detection – Sends all images + the user instruction to a Gemini function-calling model that returns a bounding-box for the most relevant region in each image.
2. Crop & Stitch – Crops each image to its ROI, then partitions and stitches them into 2–3 composite images (≈1024×1024 each) while minimising whitespace and preserving content at maximum resolution.
3. Recaption – Rewrites the user instruction so that image references (图1/图2/image N …) correctly point to the new composite images, and expands the prompt to ~512 words/characters for richer editing context. The user's original language is preserved.

(Optional) To enable the Recaption feature (rewriting instructions via Gemini for better editing results), set your Gemini API key:

export GEMINI_API_KEY="your-gemini-api-key"

Note: The Gemini API is not required. Without it, the Agent will still perform ROI detection and image stitching normally, but will skip the instruction rewriting step. Setting a Gemini API key is recommended for best results.

Training is a two-step process:

1. Extract VLM embeddings — Run offline extraction on your image–text JSONL.
2. SFT training — Train on the extracted embeddings (HSDP/FSDP, multi-node supported).

→ Full details: train/README.md (data format, environment, commands).

To better validate the capabilities of our model, we propose a benchmark called REDEdit-Bench. Our main goal is to build more diverse scenarios and editing instructions that better align with human language, enabling a more comprehensive evaluation of current editing models. We collected over 3,000 images from the internet, and after careful expert-designed selection, we constructed 1,673 bilingual (Chinese–English) editing pairs across 15 categories.

We provide the inference and evaluation code for REDEdit-Bench. Please refer to the redbench_infer.py and redbench_eval.py scripts in the tools directory for more details.

The REDEdit-Bench dataset will be available soon.

The code and the weights of FireRed-Image-Edit are licensed under Apache 2.0.

• Release FireRed-Image-Edit-1.0 model.
• Release FireRed-Image-Edit-1.0-Distilled model, a distilled version of FireRed-Image-Edit-1.0 for few-step generation.
• FireRed-Image-Edit-1.1 released with improvements to portrait consistency, multi-element fusion, stylized text reference, and portrait makeup effects.
• Release of lightweight editing models comparable to zimage/Flux2-klein and other small-parameter variants.
• Release REDEdit-Bench, a comprehensive benchmark for image editing evaluation.
• Release FireRed-Image model, a text-to-image generative model.

We kindly encourage citation of our work if you find it useful.

@article{firered2026rededit,
      title={FireRed-Image-Edit-1.0 Technical Report}, 
      author={Super Intelligence Team},
      year={2026},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.13344}, 
}

FireRed-Image-Edit has not been specifically designed or comprehensively evaluated for every possible downstream application. Users should be aware of the potential risks and ethical considerations when using this project, and should use it responsibly and in compliance with all applicable laws and regulations.

• Prohibited Use: This project must not be used to generate content that is illegal, defamatory, pornographic, harmful, or that violates the privacy, rights, or interests of individuals or organizations.
• User Responsibility: Users are solely responsible for any content generated using this project. The authors and contributors assume no responsibility or liability for any misuse of the codebase or for any consequences resulting from its use.

We would like to thank the developers of the amazing open-source projects, especially Qwen-Image for providing a powerful text-to-image foundation model, as well as Diffusers and HuggingFace.

Please contact us and join our Xiaohongshu Group if you have any questions.