Overview
Flux is an open-source AI image generation model launched by Black Forest Labs (founded by the original core team members of Stable Diffusion). Upon its release, Flux.1 shocked the AI community with its outstanding image quality and text rendering capabilities, and is considered the "spiritual successor" to Stable Diffusion, surpassing SDXL on multiple metrics.
Flux offers three versions: Flux.1 Pro (closed-source, top-tier), Flux.1 Dev (open-source, non-commercial), and Flux.1 Schnell (open-source, fast version, Apache 2.0, commercial use allowed).
The line has since expanded well beyond image generation. On 2026-07-24 Black Forest Labs launched FLUX 3, a multimodal foundation model released via Early Access that uses a unified architecture to jointly learn image, video, audio, and motion prediction, extended from the Self-Flow learning framework and capable of outputting up to 20-second videos with native audio in a single generation, with support for text-to-video, image-to-video, and multi-shot concatenation. The same day, the company detailed that video prediction accounts for over 95% of training compute and announced FLUX-mimic in collaboration with robotics company mimic. On 2026-09-23 Black Forest Labs released FLUX 3 Action, an open-weight 7B World Action Model that ranks first on the RoboLab benchmark. On 2026-10-02 FLUX 3 Image arrived on Krea and OpenRouter the same day, rendering natively up to 4K with multi-turn editing that leaves other pixels untouched, bounding-box image layout, and composition from up to 10 reference images; commercial weights are available now, with an open-weights release planned in the coming weeks.
Key Features
- Excellent Image Quality: Reaches new heights in image detail, lighting effects, and overall aesthetics, approaching Midjourney's level; FLUX 3 Image renders natively up to 4K
- Powerful Text Rendering: Significantly outperforms SD and DALL·E in rendering text within images, suitable for poster and advertisement design
- Open Source and Deployable: Dev and Schnell versions are open-source, allowing local operation and customization; FLUX 3 Action is released as an open-weight 7B model
- Multiple Model Versions: Three tiers—Pro/Dev/Schnell—balancing quality and speed needs, now extended by the FLUX 3 family across image, video, audio, and motion prediction
- High Prompt Adherence: Excellent understanding and execution accuracy for complex scene descriptions
- Rapidly Growing Community: Native support in tools like ComfyUI, with rapidly expanding LoRA and workflow ecosystem; FLUX 3 Image arrived on Krea and OpenRouter on 2026-10-02
- Multimodal Generation: FLUX 3 (2026-07-24) uses a unified architecture to jointly learn image, video, audio, and motion prediction, outputting up to 20-second videos with native audio in a single generation and supporting text-to-video, image-to-video, and multi-shot concatenation
- Advanced Image Editing and Layout: FLUX 3 Image offers multi-turn editing that leaves other pixels untouched, bounding-box image layout, and composition from up to 10 reference images
- World Action Modeling: FLUX 3 Action (2026-09-23) is an open-weight 7B World Action Model ranking first on the RoboLab benchmark
Use Cases
- Tech enthusiasts and researchers pursuing the latest open-source image models
- Designers needing to render text in images (posters, logos, etc.)
- Those seeking a free alternative close to Midjourney quality
- ComfyUI users and AI painting workflow builders
- Enterprises needing a commercial open-source model (Schnell version, Apache 2.0)
- Creators producing video with native audio, including text-to-video, image-to-video, and multi-shot concatenation
- Teams needing up-to-4K image generation with multi-turn editing, bounding-box layout, and multi-reference composition
- Researchers and developers working with open-weight world action models
Pros
- New Benchmark in Image Quality: Highest quality among open-source models, approaching Midjourney
- Leading Text Rendering: Ability to generate text in images far surpasses peers
- Schnell Version Apache 2.0: High-quality open-source model fully usable for commercial purposes
- Fast Generation Speed: Schnell version produces images in just 4 steps
- Original Stable Diffusion Team: Technical strength is guaranteed
- Multimodal Expansion: FLUX 3 jointly learns image, video, audio, and motion prediction with up to 20-second native-audio video generation
- Open-Weight World Action Model: FLUX 3 Action is a 7B model ranking first on the RoboLab benchmark
- Native 4K and Advanced Editing: FLUX 3 Image supports multi-turn editing, bounding-box layout, and composition from up to 10 reference images
Pricing
Flux.1 Schnell is completely free and open-source (Apache 2.0), usable for commercial purposes. Flux.1 Dev is open-source but limited to non-commercial use. Flux.1 Pro is accessed via API, charged per generation. Running Dev/Schnell locally is free but requires a GPU (recommended RTX 4070 12GB or higher). For the newer generation, FLUX 3 Image commercial weights are available now, with an open-weights release planned in the coming weeks, and FLUX 3 Action is released as an open-weight 7B model.
Summary
Flux remains the most noteworthy open-source image model of 2024-2025—its emergence has, for the first time, brought open-source solutions truly close to Midjourney in quality. If you already have experience with Stable Diffusion, it is highly recommended to try Flux. The Apache 2.0 license of the Schnell version makes it an excellent choice for free commercial use by enterprises. The family has since grown into a broader multimodal platform: FLUX 3 (2026-07-24) jointly learns image, video, audio, and motion prediction, FLUX 3 Action (2026-09-23) extends the line into an open-weight 7B World Action Model, and FLUX 3 Image (2026-10-02) brings native up-to-4K rendering, multi-turn editing, bounding-box layout, and up-to-10-reference-image composition to Krea and OpenRouter, with commercial weights available now and an open-weights release planned in the coming weeks.
Version History
- FLUX 3 Max launches with a statistical tie for first on AA-Video-I2V and lower pricing than FLUX 3 (2026-10-19): fal Research and Black Forest Labs co-released the post-trained video model FLUX 3 Max on October 19, scoring 1,139 on the Artificial Analysis image-to-video index to place second on AA-Video-I2V v1.1 (a statistical tie for first) and share first place with Seedance 2.5 on third-party benchmarks. A 10-second 720/768p clip takes about 12 seconds to generate, which the team says is up to 22 times faster than Seedance 2.5. Pricing sits below FLUX 3, with an additional 75% price cut running for the next three weeks.
- Black Forest Labs open-sources FLUX 3 Image Klein (5.6B) (2026-10-15): Releases FLUX 3 Image Klein, a 5.6B-parameter open-weights text-to-image model with native 4K output, pixel-perfect editing, bounding-box control, and up to 10 reference images; the INT8 build runs locally on 8GB VRAM (about 9 seconds per 1MP image on an RTX 4070); ships under the FLUX Kommunity License, free for commercial use by businesses under $5M annual revenue; API starts at $0.009 per image.
- FLUX 3 Image arrives on Krea and OpenRouter (2026-10-02): FLUX 3 Image arrived on Krea and OpenRouter the same day, rendering natively up to 4K with multi-turn editing that leaves other pixels untouched, bounding-box image layout, and composition from up to 10 reference images. Commercial weights are available now, with an open-weights release planned in the coming weeks
- Black Forest Labs releases FLUX 3 Action, an open-weight 7B World Action M (2026-09-23): Black Forest Labs releases FLUX 3 Action, an open-weight 7B World Action Model, ranking first on the RoboLab benchmark.
- inclusionAI open-sources ConceptEdit: an image editing data generation pipeline based on concept scaling and dense supervision (2026-08-17): Ant Group's inclusionAI open-sourced ConceptEdit, a data generation pipeline for image editing based on concept scaling and dense supervision. The pipeline builds a large-scale, taxonomy-based image editing dataset through a three-stage process (VLM generates instructions, FLUX performs edits, VQA evaluates and filters), and provides two variants: single-concept and multi-concept. The project is licensed under MIT, supports resuming from checkpoints, and requires an OpenAI-compatible VLM endpoint and a local FLUX checkpoint.
- Black Forest Labs releases FLUX 3 multimodal model, supporting image, video, audio, and motion prediction (2026-07-24): Black Forest Labs launches FLUX 3 multimodal foundation model via Early Access, using a unified architecture to jointly learn image, video, audio, and motion prediction. The model is extended based on the Self-Flow learning framework, capable of outputting up to 20-second videos with native audio in a single generation, supporting tasks such as text-to-video, image-to-video, and multi-shot concatenation.
- FLUX 3 x mimic: Next-generation video motion model (2026-07-24): Black Forest Labs releases multimodal foundation model FLUX 3, jointly training images, video, and audio, with video prediction accounting for over 95% of training compute. The model collaborates with robotics company mimic to launch FLUX-mimic