From FLUX.1 to the multimodal FLUX 3 — how the original Stable Diffusion team rebuilt image generation from the ground up.
When Robin Rombach, Patrick Esser, and Andreas Blattmann — the researchers behind the original Stable Diffusion architecture — left Stability AI to found Black Forest Labs (BFL) in 2024, the AI image world paid attention. Their first release, FLUX.1 (August 2024), instantly became the reference open-weight image model. Two years later, the FLUX family has grown into something far more ambitious: a full-stack media generation ecosystem spanning images, editing, video, and audio.
The FLUX Lineage
FLUX.1 (2024): The Foundation
A 12-billion-parameter rectified flow transformer shipped in three tiers:
- [schnell] — Apache 2.0 licensed, 4-step generation, runs on consumer GPUs
- [dev] — higher quality, non-commercial open weights
- [pro] — API-only flagship
FLUX.1 Kontext (2025): Editing Built In
Unlike Stable Diffusion’s bolt-on ControlNet approach, Kontext brought instruction-based image editing directly into the model — change a shirt color or swap a background with a text prompt, with strong consistency across successive edits.
FLUX.2 (November 2025): The Current Image Flagship
FLUX.2 added multi-reference conditioning (up to 10 images), 4-megapixel output, and dramatically sharper in-image text rendering. The lineup spans from API-only [pro] down to [klein], a 4B model that generates in under a second and ships preinstalled on consumer hardware. Notably, klein 4B is Apache 2.0 — the only tier you can freely embed in commercial products.
FLUX 3 (July 2026): One Model, Every Modality
The biggest architectural shift yet. FLUX 3 is a single unified network trained simultaneously on images, video, and audio — not separate models behind one interface. Key claims:
- Up to 20 seconds of 720p video with natively synchronized audio in a single pass (most competitors cap at 10 seconds and bolt audio on afterward)
- Text-to-video, image-to-video, video-to-video, and keyframe-to-video modes
- Black Forest Labs reports 93% preference wins over Luma Ray 3.2 in internal testing
- An open-weight FLUX 3 Dev release is planned for later in 2026
The same architecture is already running in the real world: FLUX-mimic, built with Zurich-based Mimic Robotics, handles soft-body manipulation on Audi’s test manufacturing lines.
Open Weights — With an Asterisk
“FLUX is open source” needs a caveat. The production-quality tiers ([pro], [flex], [max]) are closed API products. [dev] and [klein] 9B ship under the FLUX Non-Commercial License. Only [klein] 4B is Apache 2.0. If licensing is your reason for choosing FLUX, verify which variant you actually mean.
Where FLUX Wins (and Where It Doesn’t)
Wins:
- 💸 Price — FLUX.2 [pro] runs about $0.03/MP vs. $0.134 for Nano Banana Pro
- ⚡ Speed — klein models generate in under a second
- 🛠️ Self-hosting — downloadable weights (at the small end), fine-tuning ecosystems
- 🔤 Text rendering and prompt adherence
Frustrations:
- Arena Elo (~1,188) trails GPT Image 2 (~1,338) and Nano Banana 2 (~1,262) in blind human voting
- Self-hosting [dev] demands serious VRAM (64GB+ at full precision; ~19GB with Q4 quantization)
- Aggressive API filtering on IP-adjacent prompts
The Bigger Picture
Black Forest Labs has raised over $430 million and reached a $3.25 billion valuation. But the more interesting number is architectural: FLUX 3 bets that one unified foundation model beats a field of specialized ones. If video, audio, and action are all views of the same world model, the implications go far beyond prettier pictures — into robotics, simulation, and interactive media.
Final Verdict
FLUX remains the highest-quality image model family you can legally download and run on your own hardware — and with FLUX 3, Black Forest Labs is playing a longer, more ambitious game than anyone else in the image generation race. Whether a unified backbone beats specialized pipelines is 2026’s most fascinating open question.
_________________________________________________________________________________________________________________________________________________
Have you self-hosted a FLUX model, or are you waiting for FLUX 3’s open release? Let’s discuss in the comments!
Come on!