Rednote open-sources FireRed 1.1 with font cloning and locked-in faces

FireRed-Image-Edit 1.1 is an open-source semantic image editor from Rednote that lets you perform surgical edits on pictures using simple text prompts.

While version 1.0 was already capable, this new release focuses heavily on subject consistency. In plain terms, you can swap out backgrounds, modify outfits, and alter bodily poses without losing the subject's facial identity. Even challenging edge cases, like changing an infant's pose and environment, keep the facial structure fully intact.

What it can do

The model acts essentially as a text-guided photo retouching agent. You tell it what to alter in plain English, and it takes care of the adjustments.

One of its standout tricks is multi-reference fusion and virtual try-on. You can feed it multiple image elements at once, while an internal agent automatically handles image cropping and composition. Whether you want to dress a model in a dress with intricate patterns or render textured footwear, the results blend naturally without common AI warping artifacts.

Beyond basic cosmetic retouching and vintage photo restoration, it includes a clever typography feature: font style cloning. Give it a photo and a reference poster, and it can accurately extract that distinct typographical aesthetic and apply it directly onto your new composite design.

On benchmark leaderboards, FireRed 1.1 reportedly outperforms current open-source alternatives like Qwen ImageEdit and Longcat Image Edit, beating Nano Banana Pro across select evaluation metrics.

How to run it

Rednote has open-sourced the entire stack: code, model weights, technical documentation, and full LoRA training scripts. Its backbone-agnostic architecture also makes it portable across other text-to-image base models.

The real hurdle is file size. The uncompressed model sits at nearly 60GB, putting it out of reach for everyday consumer GPUs. To soften the blow, the team rolled out distilled and quantized setups alongside static compilation, trimming the VRAM requirement down to roughly 30GB and clocking end-to-end generation at 4.5 seconds.

For workflow builders, the team also included native ComfyUI custom node integration and GGUF format weights straight out of the box.

Why it matters

For digital creators and e-commerce designers, FireRed 1.1 tackles the single most frustrating bottleneck in open-source generative art: identity drift. You finally get to swap outfits and modify poses without accidentally changing your actor's face.

We think this release substantially narrows the capability gap between local workflows and commercial closed systems like Nano Banana. That said, a 30GB VRAM footprint means it remains a tool for workstations and cloud instances rather than laptop enthusiasts. The real inflection point will arrive once the open-source community manages to shave it down into lighter, consumer-friendly quants.