Black Forest Labs FLUX

innFactory AI Consulting from Rosenheim, Germany advises enterprises across the DACH region (Germany, Austria, Switzerland) on GDPR-compliant use of image generation models. FLUX by Black Forest Labs is a European AI alternative for image generation - developed in Freiburg im Breisgau, Germany by the founders of Stable Diffusion, with an EU endpoint for the API and open models for self-hosting.

German AI Innovation from Freiburg im Breisgau

Black Forest Labs is headquartered in Freiburg im Breisgau, Germany, and was founded by the creators behind Stable Diffusion. For enterprises in the DACH region, this means image generation from a German provider, with an EU endpoint and open weights that can run on your own infrastructure.

With the FLUX.2 generation (introduced in November 2025), Black Forest Labs expands its image generation offering. The family now includes five variants - [pro], [flex], [dev], [klein] (added January 2026), and the top-tier [max] (released December 2025) - covering the full spectrum from cloud API to self-hosting and on-device generation. FLUX 3, a multimodal foundation model, adds FLUX 3 Video (preview), FLUX 3 Image, and the open FLUX 3 Image Klein (details below). This overview reflects the state as of October 2026.

FLUX.2 [max] - The New Top-Tier Model

Since December 2025, FLUX.2 [max] has extended the family as the model with, per Black Forest Labs, the highest image quality and best prompt understanding. It supports up to 10 reference images simultaneously, generates images up to 4 megapixels, and, per the vendor, can incorporate current information from the web into image generation via web-based context generation - useful for subjects with a current-events angle. Billing is usage-based per megapixel (per the BFL pricing page: $0.03/MP input, $0.07 for the first output megapixel, then $0.03/MP).

FLUX.2 [pro] - The Established Flagship

The flagship FLUX.2 [pro] delivers photorealistic images up to 4 megapixels, based on a novel latent flow matching architecture combining a 24-billion parameter Mistral-3 vision-language model with a 32-billion parameter rectified flow transformer:

• 32k token context window for detailed, multi-part prompts
• Multi-reference: Up to 8-10 reference images simultaneously for consistent characters, products, and styles
• Clean typography: Precise text rendering, even at small sizes
• Photorealism: Realistic lighting and physics, eliminating the “AI look”

FLUX.2 [klein] - Open Source for Edge & Consumer Hardware

In January 2026, Black Forest Labs released FLUX.2 [klein], the fastest variant in the family:

• Sub-second generation on modern GPUs
• 4B variant under Apache 2.0 - free for commercial use
• 9B variant under Non-Commercial License
• Only ~13 GB VRAM for the 4B variant - runs on RTX 3090/4070

Current Model Overview

Superseded Models (FLUX.1)

BFL lists the FLUX.1 generation as its previous generation; it has been superseded by the FLUX.2 family:

• FLUX.1.1 Pro Ultra - superseded by FLUX.2 [pro], still bookable per the pricing page
• FLUX.1 Pro - API endpoints (flux-pro-1.0) deprecated since 31 October 2025, superseded by FLUX.2 [pro]
• FLUX.1 Dev - superseded by FLUX.2 [dev]
• FLUX.1 Schnell - superseded by FLUX.2 [klein] (4B, Apache 2.0)

FLUX 3 - Multimodal Foundation Model

With FLUX 3, Black Forest Labs announced a multimodal foundation model on 23 July 2026 that was jointly trained on image, video, and audio data. Its capabilities are being released in phases; according to the BFL documentation:

• FLUX 3 Video (preview): video with synchronized, multilingual speech (since 4 August 2026), up to 20 seconds, with modes such as text-to-video, image-to-video with keyframes, video continuation (up to 15 seconds), and draft mode. Since 10 September 2026 the video endpoint can deliver 2K (qhd) and 4K (uhd) resolution directly.
• FLUX 3 Image (since 1 October 2026): images from text, editing of specific details by instruction, up to ten reference images, and bounding boxes to place, recolour, move, or remove elements in a targeted way. Output goes up to 4K; per the release notes an image costs between USD 0.041 (768×768) and USD 0.607 (4K), depending on resolution.
• FLUX 3 Image Klein (since 15 October 2026): a compact version of FLUX 3 Image with 5.6 billion parameters and open weights on Hugging Face, with the same features. The Distilled variant is meant for fast inference, the Base variant for fine-tuning and LoRA training; per BFL the model runs on consumer GPUs down to an RTX 3090. Via the API an image costs USD 0.008 (768×768) to USD 0.036 (4K) per the release notes.
• FLUX 3 Action: a world action model (robotics, simulation, games) with 7B parameters and open weights.

The open FLUX 3 weights (Image Klein, Action) are released under the FLUX Kommunity License. Per the model card it permits non-commercial use plus commercial use of generated outputs by ‘Qualifying Users’ with less than USD 5 million in annual revenue (including affiliates). Per the model card this exception is not blanket permission for commercial fine-tuning or distribution; beyond that, BFL offers commercial licenses.

FLUX 3 runs on the same BFL API platform as FLUX.2, which offers EU and US endpoints. The BFL documentation contains no dedicated EU data residency information for FLUX 3. Via fal, fal Research also offers FLUX 3 Max, which fal describes as a variant of FLUX 3 Video post-trained together with Black Forest Labs and tuned for lower latency; it is not part of the BFL API documentation. The BFL documentation continues to list FLUX.2 as the recommended model for text-to-image; for production image generation the FLUX.2 family ([max], [pro], [flex], [dev], [klein]) therefore remains the established choice. FLUX 3 Image is worth a look for image editing with bounding boxes, FLUX 3 Image Klein as the open option for it, and FLUX 3 Video for video with sound.

Key Strengths

German Company with an EU Endpoint

• Headquartered in Freiburg im Breisgau: German company, German law
• EU Endpoints: The BFL API offers dedicated EU and US endpoints (api.eu.bfl.ai / api.us.bfl.ai); FLUX.2 [pro] additionally offers a deployment in the EU data zone (Data Zone Standard) via Microsoft Foundry
• GDPR-compliant: No data transfer to third countries required
• Made in Germany: Developed by Stable Diffusion founders

Technology

• Photorealistic Quality: FLUX.2 [pro] delivers up to 4MP with multi-reference support
• Scalable Performance: From the edge-friendly [klein] (4B) to the 32B-parameter [pro]
• Advanced Typography: Precise text rendering in generated images
• Prompt Adherence: Reliable execution of complex instructions

Open Source and Flexibility

• FLUX.2 [klein] 4B: Fully Open Source under Apache 2.0 (commercially usable)
• FLUX.2 [dev] / [klein] 9B: Open weights for research and customization
• FLUX 3 Image Klein: Open weights under the FLUX Kommunity License (commercial use of outputs below USD 5 million annual revenue)
• Fine-Tuning: LoRA and DreamBooth support for the open models
• Full Control: On-premise deployment without API dependency

Availability and Integration

Cloud Platforms

Amazon SageMaker JumpStart (not Amazon Bedrock):

• Per the AWS Bedrock model catalog, FLUX is NOT listed among the foundation models offered directly there - earlier claims about FLUX.1.1 Pro Ultra via Bedrock in eu-central-1 could not be verified anymore
• Open FLUX.2 variants (e.g. FLUX.2-klein-base-4B, FLUX.2-small-decoder) can instead be self-hosted via SageMaker JumpStart
• The region depends on the chosen SageMaker endpoint, e.g. eu-central-1 (Frankfurt) is available

Microsoft Foundry / Azure AI Foundry:

• FLUX.2 [pro], FLUX.1-Kontext-pro, and FLUX-1.1-pro can, per the region-availability table, be deployed both as Global Standard (worldwide routing) and as Data Zone Standard in the EU data zone (deployment in France Central, Germany West Central, Italy North, Poland Central, Spain Central, Sweden Central, or West Europe); a Standard/Regional deployment in a single region is not listed for FLUX there
• FLUX.2 [flex] is listed there only as Global Standard per the same table - for an EU binding without worldwide routing, [pro] is currently the right choice
• FLUX 3 is not listed on Foundry so far

Black Forest Labs API (api.bfl.ai, EU endpoint api.eu.bfl.ai):

# Direct API usage via the EU endpoint
curl -X POST "https://api.eu.bfl.ai/v1/flux-2-pro" \
  -H "x-key: $BFL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Professional product photography...",
    "width": 2048,
    "height": 2048
  }'

Additional Platforms:

• fal.ai: Fast inference API with low latency
• Replicate: Simple API integration and scaling
• Cloudflare Workers AI: FLUX.2 [dev], FLUX.2 [klein] 4B/9B globally available

Self-Hosting (On-Premise)

For maximum data control, multiple open-source options are available:

FLUX.2 [klein] 4B (Apache 2.0):

• Fully open license for commercial use
• Download via HuggingFace
• Only ~13 GB VRAM (RTX 3090/4070)
• Sub-second generation on modern GPUs

FLUX 3 Image Klein (FLUX Kommunity License):

• 5.6 billion parameters, Distilled and Base variants
• Image editing with up to ten references and bounding boxes, output up to 4K
• Per BFL runs on consumer GPUs down to an RTX 3090; the prompt upsampler needs additional GPU memory
• Commercial use of outputs only below USD 5 million annual revenue, otherwise a commercial license from BFL

FLUX.2 [dev] (Open Weights):

• 32B-parameter flagship of the open models
• FP8 quantization reduces VRAM by 40%
• Self-hosting on own servers
• Fine-tuning with LoRA and DreamBooth
• Non-Commercial License applies

Workflow Integration

• ComfyUI / Automatic1111: Native support
• Python SDK: Official integration
• REST API: Standard HTTP interface
• Batch Processing: Optimized for mass generation

Use Cases

Marketing and E-Commerce

• Product visualization in photorealistic quality
• Social media content production at scale
• A/B testing of campaign visuals
• Personalized customer communication

Design and Creativity

• Rapid prototyping of concepts
• Mood boards and storyboarding
• Iterative design with fast feedback
• Variant generation for product lines

Enterprise and Automation

• Automatic image generation for product databases
• Dynamic content for personalized websites
• Workflow integration with n8n or Make
• Scalable content pipelines with GDPR compliance

Our Recommendation

As a German AI consulting company, we especially recommend FLUX for organizations that value data sovereignty and European technology. The decisive advantage: Black Forest Labs is a German company based in Freiburg im Breisgau.

For Highest Quality Requirements:

• FLUX.2 [max] is, per Black Forest Labs, the highest-quality model in the family, with up to 10 reference images and web-based context generation
• FLUX.2 [pro] still delivers photorealistic 4MP images with multi-reference (8-10 images) and 32k-token context at a somewhat lower cost than [max]
• Both available via the BFL API (incl. EU endpoint); [pro] additionally in the EU data zone via Microsoft Foundry

For Experimental and Production-Grade Design:

• FLUX.2 [flex] offers granular control over Guidance and Steps
• Available via the BFL API (incl. EU endpoint) and Microsoft Foundry as Global Standard

For Self-Hosting and Compliance:

• FLUX.2 [klein] 4B (Apache 2.0) for commercial edge and on-device deployment
• FLUX.2 [dev] (Open Weights) for research and fine-tuning, self-hostable via Amazon SageMaker JumpStart among others
• FLUX 3 Image Klein (FLUX Kommunity License) for image editing with bounding boxes on your own hardware; commercial use of outputs only below USD 5 million annual revenue
• Full GDPR control in own infrastructure
• No data leaves your network

For Video and Multimodal Use Cases:

• FLUX 3 Video (preview) is worth a look for video with sound, FLUX 3 Image for image editing with bounding boxes; for production image generation, the FLUX.2 family remains the established choice

We are happy to advise you on optimal FLUX integration into your workflows and support technical implementation - whether cloud, hybrid, or fully on-premise.