FireRed Image Edit 1.1

Overview

FireRed Image Edit 1.1 is the second version of FireRed-Image-Edit, published by the FireRed team on 3 March 2026, less than three weeks after the 1.0 weights. It keeps the 20B-parameter design and Apache 2.0 license, and the team says it significantly improves identity consistency, multi-image conditioning and domain-specialized editing. The release notes name four focus areas: portrait consistency, multi-element fusion, stylized text reference and portrait makeup.

FireRedTeam calls 1.1 the open-source leader in character identity preservation and reports new open-source highs on ImgEdit, GEdit and REDEdit-Bench. English and Chinese prompts both work.

Choosing a variant: FireRed Image Edit runs the original 1.0 weights and returns one image per request. Both pages run the same hairstyle-grid prompt on the same portrait. The 1.0 face shifts from tile to tile, while the 1.1 grid keeps one recognisable woman in all nine. Use 1.1 for avatars, try-on and recurring characters, or when you want several candidates from one request.

What you can build

• Character sets such as avatar packs and outfit grids where one face must hold across every image.
• Virtual try-on and composites from a person photo plus garment, accessory or product shots.
• Mood and lighting changes on real photos. The harbour sample turns a hazy, pale-blue scene into a golden-hour sunset with a city skyline, while the three boats and their mooring lines stay where they were.

Inputs

• Up to five images (JPG, JPEG, PNG, GIF, WEBP or HEIC): the main photo plus references like clothing, products or a second person.
• An English or Chinese instruction. The default asks for the person's avatar with different hairstyles in a 3x3 grid.
• Advanced: samples (1 to 8 images, default 1), inference steps (1 to 50, default 30) and guidance scale (1 to 20, default 1.5).

Outputs

JPGs, as many as the samples setting asks for. Each is roughly one megapixel in the shape of the last uploaded image, so the 3:4 harbour photo comes back at 896 x 1184.

Recommended settings

Set samples to 3 or 4 to get options. Images in one request share prompt and settings and differ only in their starting noise. This page defaults to 30 steps and guidance 1.5, below the 40 steps and guidance 4 in FireRedTeam's reference script. A guidance of exactly 1 disables classifier-free guidance.

Limitations

• The team gives one to three images as the native range and, for bigger sets, crops and stitches references into two or three composites of about 1024 x 1024. Uploads on this page skip that step and go to the model as they are, so keep to three or pre-combine extras.
• The output takes the shape of the last image, so upload the photo whose framing you want to keep last.

Safety & compliance

Same Apache 2.0 license as 1.0. The project's ethics statement forbids illegal, defamatory, pornographic or harmful content and puts responsibility for generated images on the user. Identity-preserving edits make realistic images of real people easy, so edit someone's likeness only with their permission.