What is Flux.1?
Flux.1 is a family of state-of-the-art AI image generation models created by Black Forest Labs, founded by former Stability AI researchers who helped build Stable Diffusion. Flux comes in three variants: Schnell (fast, fully open-source), Dev (high quality, open-weight for non-commercial use), and Pro (highest quality, commercial API). Flux is known for exceptional prompt adherence, superior text rendering within images, natural language understanding, and high-fidelity output that rivals or exceeds SDXL and Midjourney in many benchmarks. It can run locally via ComfyUI or through cloud APIs.
Table of Contents
Flux.1 Model Variants Explained
Black Forest Labs released Flux.1 as a family of models rather than a single monolithic release, recognizing that different users have different priorities. Speed matters for rapid prototyping. Quality matters for final output. Cost matters for everyone. The three-tier approach lets users choose the right trade-off for each task.
Flux.1 Schnell
Schnell (German for "fast") is the speed-optimized variant, designed to generate images in just 1-4 inference steps. This makes it extraordinarily fast, often producing images in under 2 seconds on capable hardware. Schnell is released under the Apache 2.0 license, making it fully open-source with no restrictions on commercial use, modification, or redistribution.
While Schnell sacrifices some detail compared to Dev and Pro, its output quality is remarkably good for a distilled model. It excels at rapid iteration, batch generation, and applications where speed matters more than maximum fidelity. For many practical use cases, including social media content, concept exploration, and real-time applications, Schnell's quality-to-speed ratio is unmatched.
Flux.1 Dev
Dev is the balanced variant, offering significantly higher quality output with 20-50 inference steps. It is released as open-weight under a non-commercial license, meaning you can download, run, and modify the model freely for personal projects, research, and experimentation, but commercial use requires a separate license agreement with Black Forest Labs.
Dev is the go-to model for artists, researchers, and hobbyists who want the best possible image quality without paying per-generation API fees. It produces images with exceptional detail, coherent compositions, and remarkably accurate prompt interpretation. Text rendering, a notorious weakness in most AI image generators, is substantially better in Flux Dev than in competing models.
Flux.1 Pro
Pro is the premium variant available exclusively through API access. It delivers the highest quality output with the best prompt adherence, finest detail resolution, and most coherent complex compositions. Pro is designed for production workflows where output quality justifies the per-generation cost.
Pro is accessed through the Black Forest Labs API or through third-party platforms like Replicate, fal.ai, and Together AI. Commercial use is included in the API pricing, making it straightforward for businesses to integrate Flux into their products and workflows.
Open-Source Advantages
Flux's open-source and open-weight release strategy provides benefits that closed-source models like Midjourney and DALL-E fundamentally cannot offer. Understanding these advantages helps you decide when Flux is the right tool for your workflow.
Run Locally, Own Your Pipeline
With Flux Schnell and Dev, you can run the model entirely on your own hardware. No internet connection required, no API costs, no usage limits, no content policy restrictions beyond your own judgment. Your prompts never leave your machine, providing complete privacy. For studios working with confidential client briefs or sensitive subject matter, this local execution capability is essential.
Fine-Tuning and Customization
Open weights mean you can fine-tune Flux on your own datasets. Train a LoRA on your brand's visual style, a specific character, a product line, or an architectural style. This customization produces results no amount of prompt engineering can achieve with a closed-source model. Fine-tuned Flux models can learn your exact aesthetic preferences and reproduce them consistently.
Community Innovation
The open-source community around Flux has produced a growing ecosystem of LoRAs, ControlNet adaptors, custom nodes, and workflow templates. While smaller than Stable Diffusion's ecosystem (which has had more time to develop), the Flux community is growing rapidly and producing increasingly sophisticated tools. New LoRA models appear daily on platforms like CivitAI and HuggingFace.
Open-Source Advantages
- No per-generation costs (self-hosted)
- Complete data privacy
- Fine-tuning with custom datasets
- No content policy restrictions
- Offline operation capability
- Community LoRAs and extensions
Trade-offs to Consider
- Requires capable GPU (12-24GB VRAM)
- Setup complexity vs. web-based tools
- Self-managed updates and maintenance
- Smaller ecosystem than Stable Diffusion
- Dev license restricts commercial use
- No built-in content safety filters
ComfyUI Integration
ComfyUI has become the primary interface for running Flux locally, surpassing AUTOMATIC1111's web UI as the preferred tool among advanced users. ComfyUI's node-based workflow system provides the flexibility that Flux's architecture demands, allowing you to build custom generation pipelines with precise control over every step.
Setting Up Flux in ComfyUI
Install ComfyUI
Clone the ComfyUI repository from GitHub or use the Windows portable package. Ensure you have Python 3.10+ and PyTorch with CUDA support installed. Verify your GPU has sufficient VRAM (8GB minimum for Schnell, 12-24GB for Dev).
Download Flux Weights
Download Flux model weights from HuggingFace (black-forest-labs/FLUX.1-schnell or FLUX.1-dev). Place the safetensors files in your ComfyUI/models/unet/ directory. Also download the required CLIP and VAE models.
Load Flux Workflow
Use the built-in Flux workflow templates in ComfyUI or download community workflows. Connect the model loader, CLIP text encoder, sampler, and VAE decoder nodes in the correct pipeline order.
Configure and Generate
Set your prompt, resolution (1024x1024 recommended base), step count (4 for Schnell, 20-50 for Dev), and CFG scale (1.0-3.5 for Flux, lower than SD). Click Generate and wait for your image.
Essential ComfyUI Nodes for Flux
The Flux workflow in ComfyUI uses several specialized nodes. The UNet Loader loads the Flux model weights. The Dual CLIP Loader handles both CLIP-L and T5-XXL text encoders that Flux requires for its superior text understanding. The KSampler with Euler or DPM++ scheduling handles the denoising process. The VAE Decoder converts the latent output to a viewable image.
For advanced workflows, ComfyUI supports ControlNet integration with Flux for pose, depth, and edge-guided generation. LoRA loading nodes let you apply fine-tuned style models. IP-Adapter nodes enable image-prompted generation where you provide reference images to guide the output style or composition.
Prompt Techniques for Flux
Flux handles prompts differently than Stable Diffusion or Midjourney. Its dual text encoder architecture (CLIP-L + T5-XXL) gives it a much deeper understanding of natural language. Where Stable Diffusion works best with comma-separated keyword lists, Flux responds exceptionally well to descriptive sentences and paragraphs.
Write prompts as if you are describing an image to a talented artist. Use complete sentences, spatial relationships ("on the left side," "in the background"), and contextual details. Flux understands adjective-noun relationships, counting, color attribution, and spatial positioning far better than previous models.
Text Rendering in Images
One of Flux's standout capabilities is rendering readable text within images. Where Stable Diffusion and Midjourney struggle to produce legible text, Flux can reliably generate signs, labels, book titles, and typographic elements. Place the desired text in quotation marks within your prompt, and specify the font style, size context, and placement for best results.
Key Prompt Differences from SD
Natural Language Works Best
Write descriptive sentences instead of keyword lists. "A weathered fisherman mending nets at sunset on a wooden dock" outperforms "fisherman, nets, sunset, dock, weathered."
Lower CFG Scale
Flux works best with CFG scale between 1.0 and 3.5, much lower than SD's typical 7-12. Higher values cause oversaturation and artifacts. Start at 1.0 for Schnell, 3.0 for Dev.
No Negative Prompts (Mostly)
Flux's architecture doesn't rely on negative prompts the way SD does. Instead, describe what you want positively. Some implementations support negative prompts, but they have minimal effect.
Prompt Examples
Video Tutorials
These video guides cover Flux setup, workflow optimization, and advanced techniques.
Flux vs Stable Diffusion
Flux and Stable Diffusion share DNA (literally, given the team overlap), but they represent different generations and philosophies. Here is an honest comparison to help you decide which to use and when.
| Feature | Flux.1 Dev | Flux.1 Schnell | SDXL | SD 1.5 |
|---|---|---|---|---|
| Image Quality | Excellent | Very Good | Very Good | Good |
| Prompt Adherence | Best in class | Excellent | Good | Moderate |
| Text Rendering | Reliable | Good | Poor | Very Poor |
| Speed | 20-50 steps | 1-4 steps | 20-30 steps | 20-30 steps |
| VRAM Required | 12-24GB | 8-12GB | 8-12GB | 4-8GB |
| LoRA Ecosystem | Growing | Growing | Large | Massive |
| ControlNet Support | Available | Limited | Comprehensive | Comprehensive |
| Commercial License | Non-commercial | Apache 2.0 | Open Rail++ | Open Rail |
The practical recommendation: use Flux Dev for personal projects where you want the best possible image quality and prompt adherence. Use Flux Schnell for rapid iteration and batch processing where speed matters more than peak quality. Use Stable Diffusion when you need specific LoRAs or ControlNet workflows that have not yet been ported to Flux. Many advanced users maintain both models and choose per-task.
Pricing & Access
Flux's pricing varies significantly depending on which variant you use and how you access it. Self-hosting the open models is free beyond hardware costs, while API access to Pro follows per-generation pricing.
| Access Method | Model | Cost | Commercial Use | Notes |
|---|---|---|---|---|
| Self-Hosted | Schnell | Free (hardware only) | Yes (Apache 2.0) | Requires 8GB+ VRAM GPU |
| Self-Hosted | Dev | Free (hardware only) | Non-commercial | Requires 12-24GB VRAM GPU |
| BFL API | Pro | ~$0.05/image | Yes | Highest quality, no hardware needed |
| Replicate | All variants | ~$0.003-0.05/image | Per model license | Easy API, pay per use |
| fal.ai | All variants | ~$0.002-0.04/image | Per model license | Fast inference, serverless |
For hobbyists and artists with a capable GPU, self-hosting Flux Dev provides the best value: unlimited high-quality generations with zero marginal cost. For developers building applications, the Flux Pro API provides consistent, high-quality output without managing infrastructure. For cost-sensitive batch processing, Flux Schnell on a mid-range GPU offers the most images per dollar of any current AI image generator.
Frequently Asked Questions
Flux.1 is a family of AI image generation models created by Black Forest Labs, a company founded by former Stability AI researchers who previously worked on Stable Diffusion. Flux comes in three variants: Schnell (fast, fully open-source under Apache 2.0), Dev (high-quality, open-weight for non-commercial use), and Pro (highest quality, commercial API access). The models are known for exceptional prompt adherence, natural language understanding, and high-quality outputs that rival or surpass competing models.
Flux Schnell is the fastest model, optimized for speed with 1-4 step generation, released under Apache 2.0 for full commercial use. Flux Dev offers higher quality output with 20-50 generation steps, available under a non-commercial open-weight license for personal and research use. Flux Pro delivers the highest quality with the best prompt adherence, available exclusively through API access with commercial licensing included in pricing.
To run Flux in ComfyUI: install ComfyUI from GitHub, download the Flux model weights (Schnell or Dev) from HuggingFace, and place them in your ComfyUI models directory. You need a GPU with at least 8GB VRAM for Schnell or 12-24GB for Dev. ComfyUI's node-based interface lets you build custom workflows with ControlNet, LoRA, and other extensions. Use the Dual CLIP Loader node to handle Flux's T5-XXL and CLIP-L text encoders.
Flux generally produces better results than SDXL in prompt adherence, text rendering within images, and natural language understanding. However, Stable Diffusion has a significantly larger ecosystem of LoRAs, ControlNets, and community tools developed over a longer period. For raw output quality, Flux Dev and Pro typically outperform SDXL. For customization depth and community resources, Stable Diffusion still has the advantage. Many users use both models in complementary workflows.
Flux Schnell is fully open-source under the Apache 2.0 license, allowing unrestricted commercial use. Flux Dev is available under a non-commercial license, so commercial projects require a separate agreement with Black Forest Labs. Flux Pro includes commercial licensing in its API pricing. For commercial work, use either Schnell (free, self-hosted) or Pro (paid API) to ensure you have proper licensing.