Published August 23, 2024 · By CampaignsLive · Insights
In August 2024, a small German lab named Black Forest Labs released a family of image-generation models under the FLUX name. The team behind the lab included several of the original researchers from the Stable Diffusion project — Robin Rombach, Andreas Blattmann, Patrick Esser — who had founded Black Forest Labs after leaving Stability AI through 2023. The release was the first new foundation-model image release from the open-source side of the market that had real architectural novelty rather than incremental refinement of existing Stable Diffusion architectures.
The reception was strong. Within weeks of release, FLUX had displaced Stable Diffusion as the default open-source foundation model for several categories of professional work. The release reshaped the open-source side of the market through the end of 2024 and into 2025, and clarified several things about where the open-source ecosystem was going.
What FLUX actually delivered
Three improvements were visible immediately.
Prompt fidelity that approached the closed-source leaders. The FLUX models followed detailed prompts at a fidelity that the Stable Diffusion ecosystem had not previously achieved without significant prompt engineering work. The gap to DALL·E v3 and Midjourney v6 on prompt-following narrowed considerably.
Hands and small detail. The historical Stable Diffusion weakness on hand anatomy and small detail handling — the issue that distinguished AI-derived imagery in close inspection — was meaningfully reduced. FLUX was not perfect on hands, but it was substantially better than what the open-source ecosystem had been able to produce.
Typography and in-image text. FLUX could generate legible text within images with reliability that earlier Stable Diffusion models had not achieved. For teams producing campaign creative that needed headline integration, the improvement was practically consequential.
The release came in three model variants — FLUX.1 [pro] (closed API access), FLUX.1 [dev] (open weights, non-commercial), and FLUX.1 [schnell] (open weights, Apache 2.0) — which let the team serve different user populations with appropriate licensing. The split was unusual for the open-source side and shaped how the model was adopted in the months that followed.
How the open-source ecosystem reorganized
Three patterns became visible through the autumn of 2024.
The Stable Diffusion ecosystem of LoRAs, ControlNets, and tooling slowed in new investment but did not stop. Existing pipelines built around SDXL and the SD3 family continued to be supported and refined; new tooling investment began to flow more heavily toward FLUX-compatible adaptations. By the end of 2024, the open-source ecosystem had effectively bifurcated, with FLUX as the new default for fresh projects and Stable Diffusion as the established default for ongoing ones.
The image tooling vendors that built on top of Stable Diffusion absorbed FLUX into their stacks within months. Many of the commercial tools that used Stable Diffusion as a backend — Runway’s image tools, Krea, Leonardo, and many others — added FLUX support through Q3 and Q4 2024. Users did not, in most cases, choose which model was running underneath; the tool selected based on the prompt and parameters.
Fine-tuning workflows for brand-specific corpora adapted to the new architecture. The LoRA and full fine-tuning techniques that had matured around Stable Diffusion had to be re-implemented for FLUX. The work was straightforward but non-trivial; teams running production fine-tuning on Stable Diffusion typically took eight to twelve weeks to adapt their pipelines. By early 2025, the production-grade fine-tuning ecosystem around FLUX was approximately at parity with what had existed around Stable Diffusion before the release.
What this meant for brand teams
For brand teams running serious open-source pipelines, FLUX was both an opportunity and an integration cost. The opportunity was that the model produced meaningfully better output than the Stable Diffusion baseline that had been the working default. The integration cost was that adopting it required updating fine-tuning workflows, prompt conventions, and the surrounding production tooling.
The teams that moved fastest captured the quality improvement first; the teams that moved more slowly absorbed less integration cost but also got less of the quality lift. The trade-off was, on balance, in favor of the faster movers — the eight to twelve weeks of integration was returned within a quarter through reduced rework and tighter output quality.
For brand teams not running their own open-source pipelines, the FLUX release had a less direct effect. The commercial tools that absorbed FLUX into their backends made the quality improvement available without the integration cost. The brand-side experience was that the tools they were using got better through the second half of 2024, often without specific announcement.
Where FLUX did not change the landscape
Two things FLUX did not address.
Brand-specific output character. FLUX, like Stable Diffusion before it and like Midjourney and DALL·E in parallel, is a general-purpose image model. Its output reflects its broad training corpus. Brand creative production work that depends on the brand’s specific visual register still requires fine-tuning, reference conditioning, and the surrounding production infrastructure that handles brand consistency at the platform layer rather than at the model layer. FLUX raised the base ceiling; it did not solve the brand-specific problem.
Print and color workflow. FLUX, like its predecessors, generates in sRGB and produces output at moderate native resolution. The print-grade output workflow — CMYK with printer-specific ICC profiles, total area coverage handling, bleed and crop preparation — remains a downstream workflow rather than a native model property. For teams producing print and OOH work, this was an acceptable arrangement but it was not a one-step pipeline.
The longer significance
The FLUX release was significant beyond its immediate technical impact because it demonstrated that the open-source side of the AI image market remained healthy and could produce genuinely competitive foundation models. Through 2023 and early 2024, there had been an industry conversation about whether the closed-source platforms would simply outpace the open-source side over time, given the capital available to the closed players. FLUX answered that question in the negative — at least for the medium term — by showing that small, focused teams could still ship architectural advances that the closed leaders did not have.
The implication for brand teams choosing tools was straightforward. The open-source side was a viable production foundation. Teams that had been hesitant about committing production volume to open-source pipelines on the assumption that they would fall behind the closed alternatives could revise that assumption. The two sides of the market continued to coexist, with neither pulling decisively ahead.
For the broader history of how the major tools sat at the end of 2023, see the 2023 tool retrospective. For where the surrounding production-stack capabilities had moved by 2025, see From Generation to Production-Ready.