FLUX
FLUX is a new generation of
FLUX — Black Forest Labs’ new generation of high-quality AI image generation model
Core parameters and statistics
| Parameters | Details |
|---|---|
| Developer | Black Forest Labs GmbH (Freiburg, Germany) |
| Founders | Robin Rombach, Andreas Blattmann (formerly Stability AI / Stable Diffusion) |
| First Release Date | August 1, 2024 |
| Model Series | FLUX.1 [pro], [dev], [schnell]; FLUX1.1 [pro], [pro] Ultra |
| [schnell] License | Apache 2.0 (commercially available, fully open) |
| [dev] License | Open rights, non-commercial use |
| [pro] License | Access via API only, commercial license |
| Maximum output resolution | 4MP (FLUX1.1 [pro] Ultra) |
| API access channels | Replicate, fal.ai, Together AI, official API |
| Comparative evaluation of competing products | Image quality surpasses Midjourney v6 and DALL-E 3 in many evaluations |
The core advantage of FLUX lies in the combination of "top image quality + open strategy" - the [pro] series competes with top commercial tools such as Midjourney in terms of quality, while the [schnell] version is fully open to community and commercial projects with Apache 2.0, opening up a complete ecological chain of "open source community use → commercial deployment".
User and market recognition
After the FLUX.1 series was released in August 2024, it caused a rare sensation in the AI image generation community - the FLUX.1 [dev] model on Hugging Face became the most downloaded image generation model in the first week of release. There were a large number of review posts on Reddit and Twitter, and users generally recognized that its image quality has reached or exceeded the level of Midjourney v6.
FLUX received particularly high scores in the evaluation of the following dimensions: accuracy of human anatomy (which has historically been a difficulty in AI image generation), relationship processing of multiple objects in complex scenes, text rendering accuracy, and overall visual beauty. Multiple AI community independent reviews rank FLUX.1 [pro] and FLUX1.1 [pro] as one of the highest overall quality AI image generation models. The Stable Diffusion creative background of the Black Forest Labs team gives the product a high level of technical credibility.
Cost advantage
| Version | Access | Price | License | Applicable people |
|---|---|---|---|---|
| FLUX.1 [schnell] | Open weight download | Free | Apache 2.0 | Developers, researchers, commercial deployment |
| FLUX.1 [dev] | Open weight download | Free | Non-commercial | Research, learning, personal experimentation |
| FLUX.1 [pro] | Replicate/fal.ai API | About $0.05/card | Commercial license | Application developers, professional users |
| FLUX1.1 [pro] | API | About $0.04/piece | Commercial license | High-quality commercial application |
| FLUX1.1 [pro] Ultra | API | About $0.06/photo | Commercial license | High-resolution professional scenes |
For developers with strong self-hosting capabilities, [schnell]'s Apache 2.0 license makes it one of the lowest-cost options for high-quality commercial image generation; for users who don't want to maintain infrastructure, the pay-as-you-go model for API access is cost-effective with low to medium frequency usage.
Main functions
- High-quality text-to-image generation: The FLUX model series has reached the top level in the industry in terms of prompt word following accuracy, image realism, human anatomy accuracy and artistic style diversity, especially in the generation quality of characters and complex scenes.
- Flexible selection of three-level models: [schnell] (fastest, commercially available with Apache 2.0) meets speed priority needs; [dev] (open weight, non-commercial) is suitable for research and fine-tuning; [pro] (highest quality, API access) meets the highest quality needs, and users can flexibly choose according to actual scenarios.
- Text Rendering Capability: FLUX's accuracy in rendering text (such as poster slogans and signboard text) in images is higher than competing products of the same period. This ability gives it a practical advantage in design scenarios that require the combination of graphics and text.
- Ultra High Resolution Output (Ultra): FLUX1.1 [pro] Ultra supports image output up to 4MP, meeting scenes with extremely high resolution requirements such as printing, large-size display, and professional photography replacement.
- Open weight ecosystem ([dev]/[schnell]): Open weight allows community developers to deploy locally, fine-tune LoRA and customize styles. Communities such as Civitai have accumulated a large number of custom models and LoRA based on FLUX, enriching the diversity of generated styles.
- Fast generation ([schnell]): FLUX.1 [schnell] optimizes the inference speed and compresses the generation time while maintaining high quality. It is suitable for application scenarios that require real-time or near-real-time image generation.
- Multi-platform API access: Access is provided through multiple channels including Replicate, fal.ai, Together AI, and the Black Forest Labs official API, so developers can choose the access method that best suits their infrastructure and cost structure.
Model and version evolution
| Version | Time | Description |
|---|---|---|
| Black Forest Labs Founded | 2024-06 | Founded by Robin Rombach and Andreas Blattmann |
| FLUX.1 [pro/dev/schnell] | 2024-08-01 | The three-speed model was released simultaneously, causing a sensation in the community |
| FLUX1.1 [pro] | ~2024-10 | Double improvement in quality and speed, 3x faster |
| FLUX1.1 [pro] Ultra | ~2024-11 | 4MP ultra-high resolution version released |
| Community LoRA ecological prosperity | ~2024-Q4 | Civitai and HuggingFace communities have released a large number of FLUX LoRA |
Technical advantages
Rectified Flow Transformer Architecture: FLUX adopts a Transformer architecture (similar to DiT - Diffusion Transformer) that is different from the Stable Diffusion series (based on U-Net). Combined with the Rectified Flow training method, the model achieves significant architectural-level advantages in image detail retention, semantic understanding and generation consistency, laying a more solid foundation for subsequent quality improvements.
Guidance Distillation Technology ([schnell]): FLUX.1 [schnell] uses distillation technology to compress the generation process that requires multi-step sampling into a few-step (1-4 steps) high-quality generation, which greatly reduces the amount of inference calculations while maintaining image quality close to that of high-step sampling, achieving an efficient balance between speed and quality.
Human anatomy optimization training: The FLUX series is significantly better than the early diffusion model in terms of human anatomy accuracy (hands, facial proportions, limb joints). This comes from the team's special optimization of training data quality and supervision signals, and is one of the dimensions most praised by users in the evaluation.
How to use
| Entrance | Description |
|---|---|
| Replicate | Visit https://replicate.com/black-forest-labs for API calls or web interface trial |
| fal.ai | Visit https://fal.ai, which provides a full range of FLUX model APIs |
| HuggingFace | Download [dev]/[schnell] Open weight model, run locally or in the cloud |
| ComfyUI/Automatic1111 | Local deployment, loading FLUX model through community plug-in |
| Official API | Visit https://api.bfl.ml to directly call Black Forest Labs official API |
Typical usage steps (calling FLUX1.1 [pro] through Replicate API):
- Visit https://replicate.com to register an account and bind a payment method.
- Search "black-forest-labs/flux-1.1-pro" in Replicate to view the model page.
- Enter the prompt word experience directly in the web interface, or copy the Python sample code for API integration.
- API call: Install the
replicatePython package, set the API Token, and callreplicate.run("black-forest-labs/flux-1.1-pro", input={"prompt": "..."})to generate the image. - Local deployment ([schnell]): Download the
black-forest-labs/FLUX.1-schnellmodel weights from HuggingFace, load and run through ComfyUI (requires 24GB+ GPU memory).
Product Pricing
- FLUX.1 [schnell]: completely free, Apache 2.0 open source license, available for commercial use, local deployment, no usage limit (hardware cost is borne by yourself). Calling through the API platform is about $0.003/ticket.
- FLUX.1 [dev]: Free download with open weight, non-commercial license, suitable for research and personal learning, about $0.025/photo via API platform call.
- FLUX.1 [pro]: API access only, about $0.055/piece on Replicate, about $0.05/piece on fal.ai, suitable for commercial application development.
- FLUX1.1 [pro]: Replicate is about $0.04/piece, faster and with higher quality. It is the recommended upgraded version of [pro].
- FLUX1.1 [pro] Ultra: About $0.06/sheet, supports 4MP high resolution, suitable for professional commercial printing and design needs.
Application scenarios
1. High-quality commercial visual content generation Brand design teams and advertising companies use the FLUX1.1 [pro] API to generate high-quality product display images, marketing visuals and advertising creatives. FLUX's advantages in the quality of character, product and scene generation directly improve the commercial availability of output content and reduce post-retouching costs.
2. Developers integrate high-quality image generation Application developers can integrate AI image generation functions into their products through the FLUX API. [schnell]'s Apache 2.0 license allows commercial integration, [pro]'s high quality meets commercial application scenarios that require high-quality output, and flexible API selection to meet different cost and quality needs.
3. Research and open source community model fine-tuning AI researchers and community developers use the open weights of FLUX.1 [dev] to fine-tune LoRA and train custom models for specific styles (such as animation, realistic photography, artistic styles). The Civitai community has accumulated a large number of high-quality FLUX LoRA for users to download and use.
4. High-resolution professional image generation (Ultra) Need to be used for printing, large-size display or professional photography replacement scenes, FLUX1.1 [pro] Ultra's 4MP output meets the resolution requirements of commercial printing, combined with top-level image quality, suitable for professional visual scenes that require "AI generation + high resolution".
Applicable people
- AI Developers and Independent Developers: Need a high-quality image generation API for application integration, and [schnell]'s Apache 2.0 license enables no legal barriers to commercial use.
- AI Artists and Digital Content Creators: Pursuing top image quality, FLUX leads the way in comprehensive performance in realistic style and artistic style, and the open weight version supports local flexible control.
- Brand Design and Marketing Visual Team: Need reliable high-quality commercial image generation API, [pro] series of commercial authorization and stable API services to meet business needs.
- AI Researchers and Academics: [dev]'s open weighting supports complete research reproducibility, suitable for image generation model research, fine-tuning technology experiments, and ablation studies.
- Not suitable for scenarios: Ordinary users who have no technical knowledge at all and only need a simple drag-and-drop interface (Midjourney or Adobe Firefly provide a more friendly product experience); video generation is required (FLUX focuses on static images); local deployment that requires high GPU computing power is not feasible for ordinary consumer hardware.
Summary and Outlook
The emergence of FLUX marks another technological leap in the field of open source image generation initiated by the original team of Stable Diffusion - replacing U-Net with the Transformer architecture, challenging Midjourney with top commercial quality, while maintaining the vitality of the open source ecosystem with the [schnell] version of Apache 2.0. This dual-track strategy of "commercial version without compromising quality + open source version without compromising openness" not only ensures business sustainability, but also maintains the prosperity of the community ecology.
The current limitations are: [pro] series relies on API access, and the cost is higher than self-hosted open source solutions; running [dev]/[schnell] locally requires 24GB+ video memory, and the consumer-grade hardware threshold is high; there is no official interactive image editing interface, and it needs to rely on the UI layer of a third-party platform; the quality of Chinese prompt words still lags behind English prompt words.
Looking to the future, Black Forest Labs is expected to release a video generation model based on FLUX (continuing the technical route of Stable Video Diffusion), as well as a lighter version (lowering the threshold for local deployment of video memory); as the LoRA and ControlNet ecosystem further matures on FLUX, FLUX is expected to further narrow the gap with the Stable Diffusion ecosystem in terms of creative control flexibility.
Related tools: midjourney, stable-diffusion
Version Info
- FLUX1.1 [pro] Ultra ：Release of FLUX1.1 [pro] Ultra version, supporting ultra-high-resolution image generation up to 4MP (approximately 4 million pixels) while maintaining the top image quality of FLUX1.1 [pro], suitable for commercial design and printing purposes with professional-level requirements for output resolution, providing access through API.
- FLUX1.1 [pro] ：Released FLUX1.1 [pro], which further improves image quality, generation speed (3 times faster than the first generation) and prompt word following accuracy based on FLUX.1 [pro], competing with top commercial image generation tools in multiple independent evaluations and achieving leading scores.
- FLUX.1 series first release ([pro]/[dev]/[schnell]) ：Black Forest Labs officially releases a three-piece set of FLUX.1 models: [pro] (highest quality, API only), [dev] (open weight, non-commercial license), [schnell] (fastest, open weight, Apache 2.0 commercial license). After its release, it quickly attracted widespread attention in the AI community, and platforms such as Hugging Face, Replicate, and fal.ai were simultaneously launched for deployment.
User Reviews