FLUX.1 Fill

À propos de Black Forest Labs

Fondée 2024 · Freiburg, Germany

Black Forest Labs was founded in August 2024 in Freiburg, Germany by Robin Rombach, Patrick Esser, Andreas Blattmann and Dominik Lorenz, the team behind Latent Diffusion at LMU Munich and Stable Diffusion at Stability AI. A $31M seed round led by Andreessen Horowitz funded the launch of the FLUX.1 family (pro, dev, schnell) and the FLUX.1 Tools (Fill, Canny, Depth, Redux). The lab positions itself as the leading European open foundation-model lab for generative media.

Architecture

Rectified-flow Diffusion Transformer (DiT) with inpainting/outpainting mask conditioning

FLUX.1 Fill is the inpainting and outpainting model in the FLUX.1 Tools suite released in November 2024. It is a fine-tune of the FLUX.1 base rectified-flow DiT that takes three inputs at training and inference: the source image, a binary mask indicating which pixels should be regenerated, and a text prompt describing the desired content. The architecture concatenates the masked image latent and the binary mask as extra channels into the DiT input, similar to the design of Stable Diffusion XL Inpainting and FLUX's predecessor models, allowing the same backbone to act as a powerful editor. Outpainting is achieved by zero-padding the latent and setting the mask to 1 outside the original bounds. The model uses T5-XXL plus CLIP-L text encoders and the standard 16x VAE, sampled with 28-50 flow-matching steps. Black Forest Labs reports state-of-the-art results on prompted inpainting benchmarks at launch, outperforming Adobe Firefly and Ideogram inpainting models.

Paramètres

~12B (FLUX.1 dev backbone) plus mask-conditioning branch

Contexte

512 tokens

Capacités

• Prompted inpainting: edit a masked region with natural-language instruction
• Outpainting up to 2 MP for canvas extension
• Photorealistic blending with the surrounding context
• Strong prompt adherence via T5-XXL text encoder
• Object removal, replacement and addition workflows
• Compatible with depth/canny conditioning for hybrid pipelines
• Best for: photo editing, e-commerce retouching, advertising mockups, comic panel completion.

Entraînement et licence

Fine-tuned from FLUX.1 base weights on synthetic masked image pairs with paired captions, sampled from a large licensed image corpus.

Licence: FLUX.1 Fill [dev] is released under the FLUX.1 [dev] non-commercial license; FLUX.1 Fill [pro] is available via API for commercial use.

Tests de sécurité: Inherits safety filters of the base FLUX.1 family; additional moderation on inpainting inputs to reduce NCII and deepfake misuse.