FLUX.1 Pro

Generates high-fidelity photorealistic images from natural language prompts using a 12B-parameter flow matching architecture that enables superior prompt adherence and compositional accuracy. The model uses guidance-distilled inference to balance quality and speed across multiple variants (Pro for maximum quality, Schnell for 1-4 step inference, Dev for open-weight research). Flow matching replaces traditional diffusion schedules with continuous normalizing flows, reducing inference steps while maintaining output quality.

Generates new images by conditioning on up to 10 reference images simultaneously, enabling style transfer, compositional remixing, and multi-reference control without explicit mask-based inpainting. The model uses attention-based conditioning mechanisms (implementation details unknown) to blend visual characteristics from multiple source images while respecting text prompt constraints. Supports both photorealistic and stylized output depending on reference image selection.

Provides a web-based interface for interactive image generation, experimentation, and API key management through the Black Forest Labs dashboard. The web interface enables users to input text prompts, configure output parameters (width, height, inference steps), upload reference images, and view generated outputs. The dashboard includes a pricing calculator for estimating generation costs based on resolution and step configuration. Free tier access is available for experimentation without requiring payment. Dashboard functionality for API key management, usage tracking, and billing is implied but not detailed.

Enables user configuration of inference step count to control quality-speed tradeoff in image generation. FLUX.1 Schnell variant uses 1-4 steps for fastest inference; Pro and Dev variants support configurable step counts (exact range not documented). Inference cost scales with step count through the usage-based pricing model. More steps generally produce higher quality but slower inference; fewer steps enable faster generation with potential quality degradation. Step count is configurable through API parameters and web interface.

Provides three inference variants optimized for different quality-speed tradeoffs using guidance distillation techniques: FLUX.1 Pro (maximum quality, inference speed unknown), FLUX.1 Schnell (1-4 step inference, fastest), and FLUX.1 Dev (open-weight, guidance-distilled). Guidance distillation removes the need for classifier-free guidance at inference time by training the model to internalize guidance signals, reducing computational overhead and enabling sub-second inference on capable hardware (FLUX.2 [klein] specification). All variants share the same 12B-parameter architecture but with different training objectives and inference configurations.

Generates images with exceptional accuracy in rendering readable text, typography, and character-level details within the image composition. The model achieves this through architectural improvements in the flow matching design that better preserve fine-grained visual details compared to diffusion-based approaches. Typography rendering works across multiple languages and fonts, though language support beyond English is not explicitly documented. Text is rendered as part of the overall image generation process without separate OCR or text-specific conditioning.

Generates images with superior accuracy in human anatomy, pose, and proportional correctness compared to diffusion-based models. The flow matching architecture improves anatomical coherence through better preservation of structural relationships and spatial consistency during the generation process. Anatomical accuracy applies to full-body compositions, portraits, and complex multi-figure scenes. No explicit anatomical conditioning or pose-control parameters are documented; accuracy emerges from improved base model training and architecture.

Generates images with superior compositional accuracy, spatial relationships, and object placement consistency compared to diffusion-based models. The flow matching architecture preserves spatial coherence throughout the generation process, enabling complex multi-object scenes with correct relative positioning, scale relationships, and depth cues. Compositional accuracy applies to photorealistic scenes, technical illustrations, and abstract compositions. No explicit spatial conditioning or layout control parameters are documented; composition emerges from text prompt description and improved architectural design.

Distributes FLUX.1 Dev as open-weight model weights under the FLUX.1-dev license, enabling local deployment, fine-tuning, and research use without API dependencies. The model weights are available for download and can be run on consumer GPU hardware with sufficient VRAM. Open-weight distribution enables custom fine-tuning, integration into proprietary applications, and deployment in air-gapped or privacy-sensitive environments. Commercial use is explicitly permitted under the FLUX.1-dev license.

Provides multiple FLUX.2 model variants (klein 4B, klein 9B, flex, pro, max) optimized for different hardware and quality requirements, enabling performance scaling from edge devices to high-end inference. FLUX.2 [klein] variants are specifically optimized for local deployment with sub-second inference time on capable hardware. Parameter counts for flex, pro, and max variants are not documented. All variants share the same flow matching architecture but with different model sizes and inference configurations. The klein variants are explicitly marketed as 'ready to fine-tune' with open-weight availability.

Provides API access to FLUX.1 and FLUX.2 models through Black Forest Labs dashboard with usage-based pricing calculated by output resolution (width × height in pixels) and number of inference steps. The pricing model charges per image generated with costs scaling linearly with output dimensions. A pricing calculator is available on the website to estimate costs for different resolution and step configurations. Free tier access is available for experimentation ('Try FLUX.2 for free'). API authentication and rate limiting specifications are not documented.

Generates images at 4MP (megapixel) maximum resolution with configurable width and height parameters in pixels. Output resolution is user-selectable through API parameters or web interface, enabling optimization for different use cases (social media, print, web, etc.). The 4MP specification applies to FLUX.2 variants; FLUX.1 maximum resolution not documented. Aspect ratio flexibility is supported through independent width and height configuration. No documented constraints on minimum resolution, aspect ratio extremes, or memory requirements for different output sizes.