FLUX.1 Depth

Par Black Forest Labs

Dibināts 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 that built Latent Diffusion at LMU Munich and Stable Diffusion at Stability AI. The company raised a $31M seed round led by Andreessen Horowitz in 2024 and launched the FLUX.1 family (pro, dev, schnell) plus the FLUX.1 Tools (Fill, Canny, Depth, Redux). Black Forest Labs positions itself as Europe's flagship open foundation-model lab for generative media.

Arhitektūra

Rectified-flow Diffusion Transformer (DiT) with depth-map structural conditioning

FLUX.1 Depth is part of the FLUX.1 Tools suite released in late 2024. It extends the base FLUX.1 [dev] / [pro] rectified-flow DiT backbone with an additional conditioning branch that consumes a depth map (typically produced by MiDaS, Depth Anything v2 or ZoeDepth) of a reference image. Depth features are injected into DiT blocks via ControlNet-style adapters so that the generated image preserves the 3D layout (perspective, occlusions, scale) of the source while colour, materials, lighting and style are dictated by the text prompt. The model uses T5-XXL plus CLIP-L text encoders and a 16x VAE, and samples with 28-50 flow-matching steps. Black Forest Labs fine-tuned the conditioning branch on a large licensed corpus of depth-image pairs, building on the rectified-flow training recipe from the SD3 paper (Esser et al. 2024).

Parametri

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

Konteksts

512 marķieri

Iespējas

• Depth-conditioned generation that preserves 3D scene structure
• Works at up to 2 MP output resolution
• Strong photorealism and prompt fidelity inherited from FLUX.1
• Ideal for view-consistent restyling of indoor/outdoor scenes
• Compatible with depth maps from MiDaS, Depth Anything v2, ZoeDepth
• Pairs well with Canny and Fill for compound conditioning workflows
• Best for: interior design visualisation, automotive renders, architectural restyling, world-building.

Apmācība un licence

Fine-tuned from FLUX.1 base weights on pairs of images and estimated depth maps from a curated licensed corpus. Exact composition is not disclosed.

Licence: FLUX.1 [dev] non-commercial license for dev weights; FLUX.1 [pro] commercial only via API.

Drošības pārbaudes: Inherits safety filters of the base FLUX.1 family including CSAM, NCII and IP-related guardrails.