FLUX.1 Fill Dev

Overview

FLUX.1 Fill [dev] is Black Forest Labs' open-weights model for inpainting and outpainting. Built as a rectified flow transformer with 12 billion parameters, it takes an image, a binary mask and a text description, then fills the masked region so the new content follows the structure of the picture around it. Inpainting changes something inside the frame; outpainting extends a picture past its original borders.

Guidance distillation was part of its training. In Black Forest Labs' own inpainting benchmark it ranks second, behind FLUX.1 Fill [pro].

What you can build

Product photo apps are a natural fit: leave the item out of the mask and only its setting is repainted, while the item itself stays close to the original. Cleanup features work the same way, with a user masking a sign, a passer-by or a stain and describing what belongs there. Masking one word on a sign or label changes the wording, as the coffee banner sample shows. For a wider banner, put the picture on a larger canvas and mask the empty margin.

Inputs

• Input image: the photo you want to change.
• Input image mask: a black and white image with the same framing. White marks the area to repaint and black marks what to keep.
• Prompt: what should appear in the white area, or a description of the finished scene.

In the default input, a stainless steel double teapot stands on a kitchen counter, its mask paints the teapot black and everything else white, and the prompt reads "Change the environment with leaves."

Optional controls cover steps (25 by default), guidance scale (1 to 20, default 3.5), the number of results (1 to 8), a seed, and output width and height (up to 2048 pixels each, 1024 by default).

Outputs

You get 1 to 8 edited images at the size you set, rounded down to a multiple of 16 pixels. The model redraws the whole frame, so the area outside the mask stays close to the source but can shift slightly in color. The samples are before-and-after pairs. In one, the turban in Vermeer's Girl with a Pearl Earring becomes red hair tied in a ponytail with a thin blue headband, while the face, the earring and the dark background stay as painted. In the other, a vintage ribbon banner reading BURGER & BEER becomes COFFEE & BEER in the same engraved lettering.

Recommended settings

The card's reference code runs 50 steps at guidance 30 and outputs 1232 by 1632 pixels. This page starts at 25 steps and guidance 3.5, with guidance capped at 20, so raising both brings a run closer to that reference. Match width and height to the source photo's proportions, since the default is a square.

Limitations

The card names two problems specific to filling: colors can shift slightly in areas that were not filled, and complex textures can leave visible lines at the edges of the filled region. The card also notes that the model can miss the prompt, is sensitive to wording, may amplify societal biases and gives no factual information.

Safety & compliance

Fill [dev] ships under the FLUX.1 [dev] Non-Commercial License, which, per the card, still allows commercial use of the images it generates. Out-of-scope uses on the card include breaking the law, harming minors, spreading harmful falsehoods or personal data, and harassment. Only edit a photo of a real person if that person has agreed to it.