FireRed-Image-Edit: A General-Purpose Image Editing Model

TL;DR

FireRed-Image-Edit-1.0 is an open-weights, Apache-2.0 general-purpose image editing foundation model from FireRedTeam (Super Intelligence Team), positioned as the open-source SOTA on identity-preserving, instruction-following image editing. It is built from a text-to-image foundation rather than as a patch on an existing edit model, supports bilingual EN/ZH prompts, natively accepts 1–3 input images, and ships with an Agent module that uses a Gemini function-calling model for ROI detection + crop/stitch when the user supplies more than 3 images. The release also introduces REDEdit-Bench, a 1,673-pair bilingual editing benchmark over 15 categories, plus a forthcoming 1.0-Distilled few-step variant and a separately announced FireRed-Image text-to-image base model.

Key claims

FireRed-Image-Edit-1.0 is positioned by the authors as open-source SOTA across precise instruction following, high-fidelity generation, identity consistency, and multi-element fusion [README §Highlights].
It is “built from T2I foundations, not just a ‘patch’ on existing models” — derived from a text-to-image base (FireRed-Image, announced for separate release) rather than from a prior edit model [tweet; README §Roadmap].
The model natively supports 1–3 input images; for more than 3 images, a built-in Agent module performs ROI detection via a Gemini function-calling model, then crops and stitches inputs into 2–3 composite images (≈1024×1024 each) before editing [README §Agent].
The Agent module also provides recaption + multi-image preprocessing as a wrapper around the underlying editing model [README §Agent].
REDEdit-Bench is a new bilingual (Chinese–English) editing benchmark of 1,673 image-instruction pairs across 15 categories, constructed from ~3,000 expert-filtered web images, with inference + evaluation scripts shipped in the repo [README §REDEdit-Bench].
The announcement tweet claims 4.97/5.0 on style transfer and that the model “outperforms Nano-Banana & Seedream4.0 on GEdit benchmarks” [announcement tweet]. (No source-published GEdit table is visible in the fetched README excerpt.)
The roadmap commits to four release artifacts under Apache 2.0: the 1.0 model, REDEdit-Bench, a 1.0-Distilled few-step variant, and FireRed-Image (the underlying T2I base) [README §Roadmap, §License].
A 1.1 update is documented in the same repo and optimizes portrait consistency, multi-element fusion, stylized text reference, and portrait makeup over 1.0 [README §1.1].
The Hugging Face model card acknowledges Qwen-Image as one of the upstream open-source projects the model builds on [HF model card §Acknowledgements].

Method

The repo README does not expose a full architecture/training description; an inline citation block on the Hugging Face model card stubs out the arXiv ID (xxxx.xxxxx) for a separate technical report that ships as a PDF inside the repo (assets/FireRed_Image_Edit_1_0_Techinical_Report.pdf) rather than on arXiv at filing time. What is clearly documented is the system around the model: (a) the base is FireRed-Image, a T2I generative model trained by the same team; (b) the editing model takes 1–3 input images plus a bilingual instruction; (c) an Agent module handles >3-image inputs by issuing a Gemini function call to localize ROIs in each input, then crop-and-stitching ROIs into 2–3 composite ≈1024×1024 canvases that minimize whitespace and preserve content at maximum resolution; (d) inference is exposed via a single inference.py CLI with --input_image (N images), --prompt, --output_image, --seed. Training data, objective and architecture details belong to the bundled technical report PDF, which this page does not summarize without first reading the PDF.

Results

The fetched README does not contain a quantitative results table, and the announcement tweet’s claims should be read as marketing rather than peer-reviewed numbers: “officially the new SOTA for general image editing,” “outperforms Nano-Banana & Seedream4.0 on GEdit benchmarks,” “4.97/5.0 in style transfer” [announcement tweet]. A formal SOTA claim against closed-source frontier editors (Nano-Banana = Gemini-2.5 Flash Image; Seedream 4.0 = ByteDance) on GEdit would be a meaningful datapoint if reproduced from the technical report. The team’s own REDEdit-Bench is shipped with the release, which is useful for cross-model comparison but obviously not an external benchmark. The repo shows ≈1.2k GitHub stars and 67 forks at the time of filing.

Why it’s interesting

This is a second open-source datapoint, after HunyuanImage 3.0 Technical Report‘s HunyuanImage 3.0-Instruct, of an open-weights image editing foundation model released the same season — but at a very different point in the design space: HunyuanImage 3.0 is a native-multimodal 80B-MoE / 13B-activated unified gen+understanding system trained end-to-end on T2I + LM + MMU + interleaved + CoT data, while FireRed-Image-Edit is a single-purpose editing model derived from a smaller T2I base (built on top of the Qwen-Image lineage per the HF acknowledgements) and augmented with an external Gemini-call Agent module for multi-image preprocessing. The contrast is informative: Tencent puts the reasoning inside the model via CoT-before-pixels post-training; FireRedTeam puts it outside the model as a tool-using agent that does ROI detection and composition before the edit model runs. Both ship under Apache 2.0 with an editing-tuned variant + a base T2I + a benchmark/dataset companion, reinforcing the consolidating “open foundation release” packaging convention tracked in Open foundation-model releases. Worth tracking once the technical-report PDF is summarized.