Why FireRed-Image-Edit Is the New Powerhouse in AI Image Editing

FireRed-Image-Edit

FireRed-Image-Edit is a high‑performance diffusion‑based foundation model for image editing released by the Xiaohongshu Super Intelligence team. It achieves state‑of‑the‑art results on benchmarks such as ImgEdit and GEdit, demonstrating strong instruction understanding, identity preservation and efficient inference.

RedEdit Bench

To evaluate real‑world editing ability the authors built RedEdit Bench, a benchmark covering 15 sub‑tasks (portrait beautification, low‑resolution enhancement, object insertion, style transfer, etc.). Experiments show higher correlation with human judgments than existing benchmarks.

Data Engine

The training data are generated by a scalable engine that decomposes a complex edit into composable sub‑tasks under a “fast, controllable, precise” paradigm. Three production pathways are used:

Instruction‑controlled synthesis using expert models.

Structured control synthesis (segmentation, keypoints, depth).

Model‑agnostic template synthesis (3‑D layout, text overlays).

For long‑tail edits a “check‑and‑fill” pipeline creates targeted samples, followed by multi‑level deduplication, dozens of quality‑filter operators and strict consistency guards to ensure instruction adherence, visual naturalness and content fidelity.

Three‑Stage Training

Pre‑training : Multi‑condition bucket sampling balances task distribution; random dynamic prompts improve instruction generalisation; embedding‑based caching speeds up training.

Fine‑tuning : High‑quality curated data further boost performance.

Reinforcement learning : Asymmetric gradient optimisation combined with an OCR‑based diffusion reward penalises spelling errors, character mis‑alignments and layout collapse, improving text‑editing accuracy.

Model Capabilities

Instruction consistency

During training prompts are randomly shuffled and re‑ordered, forcing the model to learn genuine semantic‑image alignment instead of memorising fixed prompt‑image pairs.

Layout‑aware text editing

The OCR‑based reward enables accurate editing of poster text while preserving font style, size and layout, and penalises misspellings or character displacement.

Creative & multi‑image generation

The architecture supports scene synthesis from textual descriptions, style transfer, and fusion of multiple reference images, allowing high‑fidelity detail generation and multi‑reference blending.

Open‑source Release

Code, model weights and the technical report are publicly available at: https://github.com/FireRedTeam/FireRed-Image-Edit Demo space on Hugging Face:

https://huggingface.co/spaces/FireRedTeam/FireRed-Image-Edit-1.0

The team plans to continue improving portrait beautification, identity consistency and text editing, and will release updated models and the underlying base model in the coming months.