Flux

Black Forest Labs' image model with industry-leading text rendering, exceptional photorealism, and strong prompt adherence. Available in open-source and commercial variants for diverse workflows.

Free AvailableOpen SourceText RenderingPhotorealismAPI

Visit Website View Tutorial

Parameters

12B

Company

Black Forest Labs

Open Source

Schnell (Apache 2.0)

Pro Price

$0.04/image

Architecture

DiT + Flow Matching

Max Resolution

4MP (2048x2048)

Introduction

Flux represents a significant leap forward in generative AI image creation, developed by Black Forest Labs -- a team founded by researchers who created Stable Diffusion. Since its release, Flux has rapidly gained recognition for transforming text descriptions into stunning visuals that rival or exceed established players, with particular excellence in rendering clear, legible text within images -- a persistent challenge that has plagued other AI image generators.

The technical foundation of Flux is a sophisticated 12-billion-parameter hybrid architecture combining transformer and diffusion models using the DiT (Diffusion Transformer) approach. This is paired with "flow matching" methodology that enables more efficient, high-quality image generation compared to traditional diffusion techniques. The result is exceptional prompt adherence, photorealistic outputs, accurate human anatomy (especially hands and faces), and -- most notably -- the best text rendering of any AI image model.

Flux offers a tiered model family to serve different needs: Schnell for blazing-fast generation with full open-source licensing, Dev for high-quality non-commercial experimentation, Pro for professional commercial applications, and Ultra/Raw for maximum resolution and photorealism. This approach allows Black Forest Labs to foster open-source community adoption while monetizing premium capabilities, making Flux accessible to hobbyists and enterprises alike.

Pros

+Industry-best text rendering in generated images
+Excellent photorealism and human anatomy accuracy
+Strong prompt adherence and instruction following
+Free Schnell variant with full open-source commercial license
+Ultra mode for high-resolution 4MP output
+Growing LoRA and fine-tuning ecosystem
+Competitive API pricing across all tiers
+Multiple access options (web, API, local deployment)

Cons

-Full models require substantial hardware for local use
-Smaller ecosystem than Stable Diffusion (fewer community models)
-Dev model license complexity (local vs platform rules differ)
-Less artistic stylization compared to Midjourney
-Non-English text rendering less reliable
-Newer model with fewer community tutorials and resources

Key Features

Industry-Leading Text Rendering

Exceptional ability to generate clear, legible, accurately spelled text within images -- a major advancement over all previous models. Reliable for signs, logos, posters, and branded content

Strong Photorealism

Produces highly realistic images with accurate human anatomy, natural skin textures, proper lighting physics, and coherent fine details that rival professional photography

Exceptional Prompt Adherence

Accurately interprets and follows complex, detailed prompts with multiple elements. Responds well to specific instructions about composition, style, color, and spatial relationships

Schnell (Fast) Model

Apache 2.0 open-source model optimized for speed. Generates quality results in just 4 steps (seconds). Full commercial use allowed with no restrictions

Dev Model

Open-weight model offering near-Pro quality for development and experimentation. Distilled directly from the Pro model. Non-commercial locally, commercial via API platforms

Pro and Pro 1.1 Models

Commercial flagship models with highest quality, best prompt adherence, and finest details. Pro 1.1 delivers improved quality with faster generation times

Ultra Mode (4MP)

Generate images up to 2048x2048 (4 megapixels) with exceptional detail, advanced lighting effects, and accurate text rendering at high resolution

Raw Mode

Specialized mode producing authentic, photographic aesthetics. Ideal for portraits, product photography, and realistic imagery that avoids the "AI look"

LoRA Fine-tuning

Train custom styles, characters, or brand identities using 10-20 images. Available through Replicate, Together.ai, and local setups. Multiple LoRAs can be combined

FLUX.1 Tools and ControlNets

Inpainting, outpainting, redux variations, and ControlNet support (Canny edge, Depth map) for precise structural control over generated images

Who Should Use It

Text-Heavy Design and Branding

Create logos, posters, social media graphics, product mockups, and marketing materials that require clear, legible text. Flux's text rendering capability is unmatched, making it the ideal choice for any design that combines imagery with typography -- from T-shirt designs to event banners.

Graphic designers, brand managers, and marketing teams

Photorealistic Content Creation

Generate realistic product photography, stock-style images, portrait photography, and editorial content. Raw mode produces authentic photographic aesthetics, while Ultra mode delivers high-resolution output suitable for print and large-format display.

Photographers, e-commerce teams, and content creators

Custom AI Model Development

Train LoRA adaptations for specific styles, characters, or brand identities with as few as 10-20 training images. Flux's open-source ecosystem supports fine-tuning through multiple platforms, and models can be deployed via API or run locally for complete control.

AI developers, creative studios, and researchers

Local and Private Image Generation

Run Schnell or Dev models locally on your own hardware for unlimited generations with complete privacy. ComfyUI provides a node-based workflow editor for complex pipelines, while quantized versions bring the hardware requirements within reach of consumer GPUs.

Privacy-conscious users, hobbyists, and developers

Pricing Plans

FLUX.1 Schnell

$0/forever

Apache 2.0 open-source license
4-step fast generation (seconds)
Full commercial use allowed
Local or API deployment options
Good quality at very high speed
Community LoRA support

Recommended

FLUX.1 Dev

$0 local / ~$0.025 API/per image via API

Non-commercial local; commercial via platforms

Open weights on Hugging Face
Near-Pro quality output
Non-commercial license for local use
Commercial via Replicate/Fal.ai APIs
Great for development and prototyping
LoRA training support

FLUX 1.1 Pro

$0.04/per image

Via BFL API or partner platforms

Highest quality output available
Best prompt adherence and detail
Full commercial license included
Faster generation than original Pro
Access via multiple API partners
Enterprise-ready reliability

FLUX 1.1 Pro Ultra

$0.06/per image

High-resolution mode up to 4MP

Up to 4MP resolution (2048x2048)
Exceptional fine detail and texture
Advanced lighting and atmosphere
~10 seconds per image generation
Text rendering at high resolution
Commercial license included

Web Platforms

$10.90-25.90/monthly subscription

Flux1.ai, FluxPro.ai, getimg.ai, etc.

No technical setup required
User-friendly web interface
Multiple Flux model access
Commercial license included
Free tiers or trials available
Credit-based billing systems

How It Compares

Flux vs Stable Diffusion

Flux and Stable Diffusion are both available for local use, but serve different strengths. Flux offers significantly better output quality, text rendering, and prompt adherence out of the box. Stable Diffusion has a much larger ecosystem of community models, LoRAs, and extensions, plus lower hardware requirements for older versions.

Flux wins at

+Much better text rendering in generated images
+Higher baseline quality without extensive tuning
+Superior prompt adherence and photorealism
+More efficient architecture with flow matching

Stable Diffusion wins at

+Stable Diffusion has a vastly larger model ecosystem (thousands of models)
+SD 1.5 runs on much lower-end hardware (6GB VRAM)
+Stable Diffusion has more ControlNet variants and extensions
+Larger community with more tutorials and resources

Flux vs Midjourney

Flux and Midjourney target different creative needs. Midjourney produces the most aesthetically pleasing, artistic images with superior composition and mood. Flux excels at technical accuracy -- text rendering, photorealism, prompt adherence, and anatomical correctness. Midjourney is subscription-only; Flux offers free open-source options.

Flux wins at

+Far superior text rendering in images
+Open-source model available for free local use
+Better photorealism and anatomical accuracy
+Flexible per-image API pricing vs subscription

Midjourney wins at

+Midjourney has superior artistic quality and aesthetics
+Midjourney offers Style and Character References for consistency
+Midjourney has a more polished user experience
+Midjourney has a larger creative community

1. Getting Started (Web Platforms)

The easiest way to use Flux is through web interfaces that require no technical setup: **Flux1.ai / FluxPro.ai:** 1. Visit the site and create an account 2. Get free credits to start experimenting 3. Enter your text prompt describing the image you want 4. Select your model (Schnell for speed, Dev for quality, Pro for best results) 5. Choose aspect ratio and any additional settings 6. Click Generate and download your images **getimg.ai:** - Offers 100 free images per month - Access Schnell, Dev, and Ultra in Essential mode - Clean interface with batch processing support These platforms handle all technical complexity, making Flux accessible to everyone regardless of technical background.

2. Using Flux via API

For developers and power users, API access offers more control and integration possibilities: **Replicate:** ```python import replicate output = replicate.run( "black-forest-labs/flux-schnell", input={"prompt": "A cyberpunk cityscape at night with neon signs reading 'OPEN 24/7'"} ) ``` **Together.ai, Fal.ai, and BFL direct API** also offer Flux access with OpenAI SDK compatibility in many cases. **Pricing comparison per image:** - Schnell: ~$0.003 (essentially free) - Dev: ~$0.025 - Pro 1.1: ~$0.04 - Ultra: ~$0.06 For high-volume use, API pricing is often more cost-effective than subscription-based platforms. Context caching on some platforms reduces costs for repeated prompt prefixes.

3. Running Flux Locally (ComfyUI)

**Hardware Requirements:** - 12GB+ VRAM recommended for full quality (RTX 4070 Ti or better) - 8GB VRAM possible with FP8 or NF4 quantization (some quality loss) - 24GB+ VRAM ideal for full models without compromises **Setup in ComfyUI:** 1. Update ComfyUI to the latest version 2. Download model files from Hugging Face: - UNET: flux1-schnell.safetensors (or flux1-dev.safetensors) - VAE: ae.safetensors - CLIP encoders: clip_l.safetensors + t5xxl_fp8_e4m3fn.safetensors 3. Place files in the appropriate ComfyUI model directories 4. Load a pre-made Flux workflow JSON from the community **For lower VRAM (8-12GB):** - Use FP8 or GGUF quantized model versions - Enable model offloading to system RAM - Consider Forge UI for better memory efficiency - Use Schnell (4 steps) instead of Dev (20+ steps)

4. LoRA Training for Custom Styles

Train custom styles, characters, or brand identities: **Via Replicate (easiest):** 1. Prepare 10-20 high-quality, consistent training images 2. Use the flux-dev-lora-trainer on Replicate 3. Training typically costs ~$1.85 and takes 15-30 minutes 4. Receive LoRA weights file for immediate use **Via Together.ai:** 1. Upload your training dataset 2. Configure training parameters (epochs, learning rate) 3. Pay per megapixel pricing ($0.035/MP) **Local Training:** Use community Kohya-style trainers adapted for Flux architecture **Using trained LoRAs:** - Add your trigger word to the prompt - Adjust LoRA strength (0.5-1.0 is typical) - Multiple LoRAs can be combined for complex effects - Works in ComfyUI, Automatic1111/Forge, and via API

Frequently Asked Questions

Flux excels at text rendering (significantly better than both), photorealism, and prompt adherence. Midjourney produces more artistic and stylized results with superior composition. Stable Diffusion has a much larger model ecosystem and lower hardware requirements. Many creators use multiple tools for different needs.

Yes. Schnell is Apache 2.0 licensed for full commercial use with no restrictions. Pro and Ultra models include commercial licenses when accessed via paid APIs. Dev is non-commercial when run locally, but commercial when generated via platforms like Replicate -- always verify specific platform terms.

Full models work best with 24GB+ VRAM (RTX 4090, A100). Optimized versions (FP8, GGUF, NF4 quantization) can run on 12GB consumer GPUs like the RTX 4070 Ti. 8GB is possible with heavy quantization and some quality tradeoffs. For most casual users, API access is more practical.

Schnell: Fastest (4 steps), open source, good quality, free. Dev: Higher quality, distilled from Pro, non-commercial locally. Pro/Pro 1.1: Best quality and detail, commercial, closed-source. Ultra: 4MP high resolution. Raw: Optimized for authentic photographic aesthetic.

Flux has the best text rendering of any AI image model, significantly better than Stable Diffusion, Midjourney, or DALL-E. It can reliably generate legible English text on signs, posters, logos, and product labels. Non-Latin scripts and very long text may be less reliable.

Flux Pro (~$0.04/image) is very competitive. Schnell is completely free for local use under Apache 2.0. Compared to Midjourney subscriptions ($10-120/month), Flux API is cheaper for high-volume use. Web platform subscriptions ($10-25/month) offer predictable monthly costs.

Yes. LoRA training is available through Replicate ($1-2 per training run), Together.ai, and local setups with community training scripts. You need 10-20 high-quality training images. Multiple LoRAs can be combined during generation for complex effects.

Flow matching is the core generation technique Flux uses instead of traditional diffusion denoising. Rather than iteratively removing noise step by step, it learns direct transformation paths between distributions, resulting in faster, more efficient, and higher-quality image generation.

Video generation capabilities are emerging but not yet a primary feature. Some community implementations exist for short video clips, but Flux is primarily an image generation model. For AI video, consider dedicated tools like Runway, Kling, or Sora.

Flux offers significantly better text rendering, superior photorealism, and more flexible deployment options (open source, API, local). DALL-E 3 is more accessible through ChatGPT and better at following complex conversational instructions. Both produce high-quality images but serve different workflows.