Technical creative teams
Run an adaptable image editor on controlled infrastructure and connect it to an internal production pipeline.
Independent tool overview
FireRed-Image-Edit at a glance
FireRed Image Edit is an Apache-licensed, 20-billion-parameter image editing model from Xiaohongshu's FireRed Team. Version 1.1 focuses on identity consistency, multi-element fusion, portrait work, stylized text, and local deployment through Diffusers or ComfyUI.
Visit the official FireRed-Image-Edit site ↗
- Developer
- FireRed Team at Xiaohongshu
- Current version
- FireRed Image Edit 1.1
- Model size
- 20B parameters
- License
- Apache 2.0
- Native inputs
- 1-3 images
- Optimized VRAM
- About 30GB
Overview
What FireRed-Image-Edit is
FireRed Image Edit is an open-weight model for changing existing images from natural-language instructions. It can add, remove, replace, restyle, restore, retouch, and combine visual elements while trying to preserve the parts of the source image that should remain consistent.
The current 1.1 release builds on the 1.0 technical report with improvements to portrait identity, multi-image conditioning, stylized-text reference, makeup, and other production-oriented edits. The official model card lists 20 billion parameters and BF16 weights.
This is primarily a model and developer toolkit, not a polished hosted design subscription. The code and weights are free under Apache 2.0, but practical use requires capable GPU infrastructure, a supported local workflow, or a third-party host.
Use cases
Who FireRed-Image-Edit is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Portrait and fashion workflows
Preserve recognizable subjects while changing makeup, clothing, styling, backgrounds, or other visual details.
Multi-product compositions
Fuse people, apparel, props, products, and scenes from multiple source images into one directed composition.
Model researchers and trainers
Study the released benchmark and training code or build specialized LoRA adapters for repeatable domain-specific edits.
Capabilities
Core FireRed-Image-Edit features
Instruction-based editing
Make localized or broad edits with text instructions, including additions, removals, replacements, styling, retouching, and restoration.
Identity consistency
Version 1.1 emphasizes keeping a person recognizable across changes to the surrounding scene, wardrobe, makeup, and composition.
Multi-element fusion
Natively condition on one to three images, or use the companion Agent workflow to crop and stitch inputs when a composition contains more sources.
Text-style reference
Carry typography and stylized text treatments from a reference into an edited output for posters and branded creative work.
Local deployment options
Run with Diffusers, the repository's inference script, a native ComfyUI workflow, or community GGUF quantizations.
Open training stack
Use the released training code, offline feature extraction, distributed-training support, and LoRA ecosystem to adapt the model.
Process
How the FireRed-Image-Edit workflow works
Step 1
Choose a model build
Start with version 1.1 for the latest quality improvements, or evaluate the distilled and quantized options when speed or memory is the constraint.
Step 2
Prepare the environment
Install the repository dependencies or a compatible Diffusers and ComfyUI stack, then download the official weights from Hugging Face or ModelScope.
Step 3
Load source images
Provide one to three inputs directly. For larger multi-image jobs, configure the Agent preprocessing path that identifies regions, crops them, and stitches composite inputs.
Step 4
Write a precise instruction
Identify the element to change, describe the intended result, and state which faces, products, poses, typography, lighting, or background details must stay fixed.
Step 5
Generate and compare
Test seeds and model variants, inspect identity and composition, and compare the optimized path with the full model before standardizing a workflow.
Step 6
Apply publication checks
Review anatomy, text, likeness, product accuracy, manipulation disclosure, source-image rights, and privacy before distributing an edited image.
Cost
FireRed-Image-Edit pricing and free plan
FireRed Image Edit's official code and model weights are free under Apache 2.0. There is no required FireRed subscription. The real cost is GPU hardware, cloud compute, storage, engineering time, and any optional language-model API used for Agent instruction rewriting.
Code and weights
Free
Download, modify, and self-host the official project under the Apache 2.0 license.
- Version 1.1, 1.0, and distilled weights are published
- Commercial use is permitted subject to the license
- Infrastructure is not included
Self-hosted inference
Compute costs vary
Run the full or optimized model on your own workstation or cloud GPU.
- Official optimized example targets about 30GB VRAM
- The team reports about 4.5 seconds per sample on its optimized setup
- Actual speed and cost depend on GPU, resolution, and configuration
Agent recaptioning
Optional API cost
Basic use does not require an LLM API, but automatic instruction rewriting can call a configured provider.
- Gemini is the default recaption option
- MiniMax and OpenAI-compatible endpoints are supported
- ROI detection for the Agent's multi-image path uses Gemini
Pricing checked . Check current pricing at the source ↗
Assessment
FireRed-Image-Edit strengths and limitations
Where it stands out
- Apache 2.0 code and weights with no required model subscription
- Strong focus on identity preservation and high-fidelity image editing
- Version 1.1 supports portraits, makeup, stylized text, restoration, and multi-element compositions
- Official Diffusers, training, optimized inference, and ComfyUI paths
- Agent preprocessing extends complex compositions beyond the native one-to-three-image input
- Published technical report, benchmark code, and model weights make the project inspectable
What to consider
- The 20B BF16 model is demanding; even the team's optimized example requires roughly 30GB of VRAM
- There is no official managed inference provider listed on the Hugging Face model page, so deployment is primarily self-hosted or community-hosted
- The Agent's larger multi-image workflow adds preprocessing complexity and uses Gemini for ROI detection
- Headline benchmark results come from the model creators and include the team's own REDEdit-Bench
- Image edits involving real people, copyrighted assets, or brand materials need explicit rights, privacy, and disclosure controls
- It is an image editor rather than a complete browser-based design application
Compare
FireRed-Image-Edit alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Content Creator
FLUX.2
Choose FLUX.2 when you want another open visual model family with image generation, editing, and a broader commercial API ecosystem.
Explore FLUX.2 →Content Creator
Seedream 4.5
Choose Seedream 4.5 for a managed usage-based API with 4K output and up to 10 direct reference images.
Explore Seedream 4.5 →Content Creator
Nano Banana 2
Choose Google's managed model when ease of access, fast generation, and Gemini integration matter more than self-hosting.
Explore Nano Banana 2 →Questions
FireRed-Image-Edit FAQs
What is FireRed Image Edit?
FireRed Image Edit is an open-weight diffusion model for modifying existing images from text instructions. It supports general editing, portrait work, multi-image composition, stylized text, restoration, and specialized adapters.
Is FireRed Image Edit free?
Yes. The official code and weights are released under Apache 2.0. You still pay for the GPU, storage, engineering, and any optional third-party API used in your deployment.
How much VRAM does FireRed Image Edit need?
The FireRed Team says its optimized inference path can run with about 30GB of VRAM and produce a sample in roughly 4.5 seconds. Requirements and speed vary with hardware, model build, resolution, and optimization settings.
Does FireRed Image Edit work in ComfyUI?
Yes. The team publishes native ComfyUI packages and workflows for the model, alongside Diffusers and command-line inference options.
What changed in FireRed Image Edit 1.1?
Version 1.1 improves portrait identity consistency, multi-image conditioning and fusion, stylized-text reference, portrait makeup, and other domain-specific editing tasks over the 1.0 foundation model.
Bottom line
Our FireRed-Image-Edit verdict
FireRed Image Edit is compelling for technical teams that want a powerful, modifiable image editor with open code, open weights, strong identity preservation, and a serious training and deployment stack. Its main tradeoff is infrastructure: creators seeking a simple hosted editor will be better served by a managed alternative.
Visit FireRed-Image-Edit website ↗Similar Design Tools
View all Design AI tools →
Qwen Image 2.0
Qwen Image 2.0 - Alibaba's new unified image generation and editing model with native 2K resolution and 1K-token text rendering
Recraft V4
Recraft V4 - New image model tuned for design taste, typography, and production-ready outputs
Riverflow 2.0
Riverflow 2.0 - Sourceful's new SOTA image editing model
Seedream 5.0 Lite
Seedream 5.0 Lite - ByteDance's updated text-to-image model with style control and real-time trend generation
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.