FLUX.1 Canny

About Black Forest Labs

Founded 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 core team that previously developed Latent Diffusion Models (LDM) and Stable Diffusion at LMU Munich and Stability AI. The team's earlier work includes the seminal 'High-Resolution Image Synthesis with Latent Diffusion Models' (CVPR 2022) and the Stable Diffusion 1.x/2.x and SDXL releases. The company raised a $31M seed round led by Andreessen Horowitz in 2024 and launched the FLUX.1 model family (pro, dev, schnell) along with specialist Tools models (Fill, Canny, Depth, Redux). Their stated goal is to be the leading open foundation-model lab for generative media in Europe.

Architecture

Rectified-flow Transformer (DiT) with Canny-edge structural conditioning

FLUX.1 Canny is part of the FLUX.1 Tools family released by Black Forest Labs in late 2024. It extends the base FLUX.1 [dev] / [pro] rectified-flow transformer (a Diffusion Transformer trained with the flow-matching objective rather than classical DDPM) with a structural conditioning branch that ingests a Canny edge map of a reference image. The conditioning is implemented as a ControlNet-like adapter that injects edge features into the DiT blocks via cross-attention residuals, so that the output respects the line structure of the input while the prompt controls semantics, style and lighting. The backbone uses T5-XXL plus CLIP-L as text encoders and operates in a learned latent space at 16x downsampling. Sampling typically uses 28-50 flow-matching steps with the Euler or DPM-Solver sampler. The model was distilled and fine-tuned on edge-conditioned pairs derived from a large licensed image corpus.

Parameters

~12B (FLUX.1 dev backbone) plus ControlNet-style conditioning

Context

512 tokens

Capabilities

• Edge-conditioned image generation that preserves line structure of a reference
• Works at up to 2 MP output resolution
• Strong photorealism inherited from FLUX.1 [dev]/[pro]
• Excellent prompt adherence via T5-XXL text encoder
• Useful for redesigns, restyles and product re-skins on a fixed silhouette
• Compatible with the rest of the FLUX Tools family (Fill, Depth, Redux)
• Best for: e-commerce restyling, architectural restyling, fashion mockups, design iteration.

Training & license

Fine-tuned from FLUX.1 base weights on pairs of images with extracted Canny edge maps. Exact dataset is not disclosed; Black Forest Labs states licensed and curated data.

License: FLUX.1 [dev] non-commercial license for the dev weights; FLUX.1 [pro] available only via API. Commercial use requires the pro/enterprise tier.

Safety testing: Standard Black Forest Labs safety filters; the model inherits guardrails of the base FLUX.1 family including filters for CSAM, non-consensual intimate imagery and IP-protected content.