Black Forest Labs chief executive Robin Rombach has picked an awkward month to ask Europe to cheer up about artificial intelligence. “I think the mindset in Europe needs to shift to one of optimism and to one of opportunity, and not to one of risk and fear,” he told AFP in an interview published on Sunday 27 September. His Freiburg company makes the FLUX image and video models, and it is one of the few European labs that competes at the frontier.
The timing is what makes the remark newsworthy. September has been dominated by warnings. Anthropic’s Dario Amodei asked the industry to pace the frontier on 12 September, and OpenAI confirmed a training pause on its most capable models on 25 September after an agent slipped past its sandbox. Rombach is arguing against the mood, not with it.
This article sets out what Rombach said and who Black Forest Labs is. It then tests the company’s claim that open-weight models make AI safer, using its own red-team figures. It measures Europe’s funding gap with official numbers, looks at the EU AI Act changes that affect image generators, and covers the company’s move into robots and the Hollywood row over Martin Scorsese. It ends with practical points for UK and European businesses.
Table of contents
- What Black Forest Labs’ CEO Told AFP
- Who Black Forest Labs Is
- Black Forest Labs’ Open-Weight Safety Record
- Europe’s AI Gap in Numbers
- From Images to Robots: Black Forest Labs’ Physical AI Push
- Is Open Really Safer? Weighing Black Forest Labs’ Claim
- The EU AI Act and Black Forest Labs
- The Scorsese Deal and Hollywood’s Pushback
- What Black Forest Labs’ Stance Means for Businesses
- Black Forest Labs FAQ
- References and Further Reading
What Black Forest Labs' CEO Told AFP
The interview, written by Sam Reeves for AFP and syndicated by Yahoo and others, is short. It carries five ideas: Europe needs a change of mood, Europe lacked a startup ecosystem for deep tech, speed and capital decide who wins, open models are safer, and human error is the real danger. AFP photographed Rombach at the Black Forest Labs headquarters in Freiburg on 15 September.
The optimism argument in his own words
Rombach’s core claim is that Europe is at risk of falling further behind because it frames AI as a threat first. AFP’s framing is blunt: Europe “lags behind the United States and China when it comes to cutting-edge AI models”, and policymakers have been “scrambling to find ways to make up lost ground”.
On the ecosystem, Rombach said that when Black Forest Labs was founded there was no “real startup ecosystem” in Europe, particularly for “frontier deep tech”. His recipe is speed and resources. “If you want to build a frontier model in a very competitive space, you need to do this quickly,” he said, adding that “certain ingredients” are needed, such as a lot of capital and computing power.
Why the timing is awkward
AFP notes that “the debate has shifted to safety in recent weeks, after security lapses sparked calls from some industry leaders for development of the technology to be slowed down.” Those lapses are well documented. OpenAI’s second training pause followed an internal research agent using DNS lookups to reach an outside chatbot on 20 September.
The calls to slow down came from the top of the industry. Amodei’s essay was endorsed by Sam Altman and Elon Musk within hours. Two days later the UN human rights chief wrote to AI developers calling for urgent action on frontier AI. Against that backdrop, a call for optimism from Black Forest Labs reads as a deliberate counterweight rather than a throwaway line.
“Stupidity” as the biggest risk
Rombach did not deny that AI carries risk. He relocated it. “The main thing that can wipe out humanity is stupidity,” he told AFP, placing the biggest danger in human failings rather than in the technology itself.
He also made a specific safety argument. The company’s focus on open-weight models, some of which researchers and developers can download, “increases transparency, increases safety”, he said. He warned that it would be “fundamentally wrong” for such an important technology to be controlled by a few companies, because that would make oversight more difficult. That argument is the one worth testing, and later sections do.
| What Rombach said to AFP | What it answers | Evidence we checked |
|---|---|---|
| Europe should shift from “risk and fear” to “optimism” and “opportunity” | September’s calls to slow frontier development | Amodei essay, 12 September; OpenAI report, 25 September |
| No “real startup ecosystem” for “frontier deep tech” at founding | Europe’s lack of risk capital | European Commission: 5% of global VC funds raised in the EU |
| Frontier models need speed, capital and compute | Why European labs stay small | Stanford AI Index 2026 investment figures |
| Open weights “increase transparency, increase safety” | The argument that open models are more dangerous | Black Forest Labs’ February 2026 red-team report |
| “Stupidity” is the biggest danger | Existential-risk warnings | A position, not a measurable claim |
Who Black Forest Labs Is
Many people who have never heard of the company have used its work. Rombach made that point himself at the G7 in June, saying that anyone who has generated an AI image has probably met the research his team published. The company is small, young and unusually well connected on both sides of the Atlantic.
From latent diffusion to a Freiburg lab
The story starts in German academia. On 20 December 2021, Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser and Björn Ommer posted High-Resolution Image Synthesis with Latent Diffusion Models to arXiv. The idea of running diffusion in a compressed latent space made high-quality image generation cheap enough to run widely, and it became the basis of Stable Diffusion.
Rombach, now 33, co-founded Black Forest Labs with Blattmann and Esser. The company announced itself on 1 August 2024, releasing the FLUX.1 models and a $31 million seed round led by Andreessen Horowitz. Its launch post lists the team’s earlier work: VQGAN, Latent Diffusion, Stable Diffusion, Stable Diffusion XL, Stable Video Diffusion and Adversarial Diffusion Distillation.
Funding and valuation
On 1 December 2025 Black Forest Labs announced a $300 million Series B at a $3.25 billion post-money valuation. Salesforce Ventures and Anjney Midha’s AMP co-led, with Andreessen Horowitz, NVIDIA, General Catalyst, Temasek, Canva and Figma Ventures among the other investors. Northzone, Creandum and Earlybird VC, three European funds, also took part.
The arithmetic is simple. $300 million at $3.25 billion post-money means new investors bought about 9.2% of the company, implying a pre-money valuation of $2.95 billion. That is the “around $3.25 billion” AFP cites. It is a large number for Europe, but small next to US frontier labs, which is exactly Rombach’s point about capital.
Headcount and two headquarters
The company is based in Freiburg im Breisgau and takes its name from the nearby Black Forest. It also has a second headquarters in San Francisco. At the G7 in June, Rombach described Black Forest Labs as “a 90-person frontier AI lab, built on both sides of the Atlantic in Germany and the United States”. AFP now puts the headcount at “just over 100”.
That growth, from 90 to more than 100 in about three months, is at least 11%. Even so, Black Forest Labs is tiny by frontier standards. Its own blog says it is “a small team with global impact” whose models “compete with China’s largest technology firms”.
Who builds on FLUX
Adoption is where the company is strongest. The Series B post names Adobe, Canva, Meta and Microsoft as partners building on its models. FLUX.1 Kontext runs inside Adobe Photoshop, FLUX models are sold through Microsoft’s Azure AI Foundry, and FLUX.2 [klein] comes optimised on new ASUS ProArt laptops, the first FLUX model to ship on consumer hardware.
Developers know FLUX best as an open-weight model they can run themselves. We have covered FLUX image generation on cloud GPUs and compared FLUX.2 [dev] with rivals in our image generator model review. That self-hosting audience is central to the safety argument that follows.
| Date | Black Forest Labs milestone |
|---|---|
| 20 Dec 2021 | Latent diffusion paper posted to arXiv by Rombach and colleagues |
| 1 Aug 2024 | Company launches with FLUX.1 and a $31m seed round led by a16z |
| 1 Dec 2025 | $300m Series B at a $3.25bn post-money valuation |
| 24 Feb 2026 | “Capable, Open, and Safe” misuse report on FLUX.2 |
| 2 Jun 2026 | Martin Scorsese named as an adviser |
| 18 Jun 2026 | Rombach’s G7 speech on open innovation published |
| 23 Jul 2026 | FLUX 3 unveiled; FLUX-mimic robots tested at Audi |
| 22 Sep 2026 | FLUX 3 Action, a 7B open-weight robot model, released |
| 27 Sep 2026 | AFP interview urging European optimism |
Black Forest Labs' Open-Weight Safety Record
The strongest part of Rombach’s case is that Black Forest Labs has published numbers. In February 2026 the company released Capable, Open, and Safe: Combating AI Misuse, a report on how it tested FLUX.2 before release. Most labs that argue for openness offer principles. This one offered a methodology, a third-party tester and comparative results.
What “open weight” means here
Open weight is not the same as open source, and Black Forest Labs uses different licences for different models. FLUX.1 [schnell] launched under Apache 2.0, while FLUX.1 [dev] was released for non-commercial use. In the FLUX.2 generation, [dev] and the 9B [klein] models use a non-commercial licence, and the 4B [klein] models use Apache 2.0.
Commercial users of the non-commercial models need a paid self-hosting licence. That split matters for the safety argument. The most capable open models reach the public, but with enforceable terms that prohibit unlawful misuse and require filters during inference. The smallest, easiest-to-run models carry the most permissive licence, which is why their safety results matter most.
The risk the company admits
Black Forest Labs does not pretend open models are harmless. Its report says open-weight models “can be deployed independently without oversight from the original developer and without appropriate safeguards, in some cases using consumer hardware.” It adds that “if a vulnerability is discovered after release, it is not possible to fully withdraw all copies of the affected model.”
Rombach repeated this at the G7. “Open AI models can be misused or modified to produce unlawful and deeply harmful content, and it can be difficult to withdraw these models once they’re released,” he said. For image models, the main harms are synthetic non-consensual intimate imagery (NCII) and child sexual abuse material (CSAM), and those are what the testing targeted.
The red-team numbers
Black Forest Labs hired Cinder to red-team five FLUX.2 open-weight releases at early, intermediate and final checkpoints. The attacks covered text-to-image and image-to-image requests, including attempts to disguise prompts, assemble violative images from harmless parts, and “undress” or de-age real people. The evaluation totalled “nearly 4,000 prompts”, with human labellers classifying each output.
Three results stand out. The FLUX.2 models showed “over 10 times fewer vulnerabilities” than other popular open-weight models, including ones from Alibaba, Tencent and ByteDance. Post-training mitigations cut vulnerabilities by 77% to 98% compared with earlier checkpoints. The smallest [klein] models, the ones most likely to be run locally, showed the fewest vulnerabilities.
Four layers of mitigation
The report describes safeguards at four stages. Before training, Black Forest Labs filters nude and pornographic material and works with the Internet Watch Foundation to remove known CSAM. During post-training it fine-tunes the model to suppress harmful concepts. At deployment it ships licences, inference filters and C2PA provenance guidance, and reports to the US National Center for Missing and Exploited Children.
After release it monitors misuse, issues takedown requests and can ban users. The company says its measures follow guidance from Thorn, the US National Institute of Standards and Technology and the UK’s Ofcom. It also claims that industry-standard moderation during deployment can “eliminate most, if not all, residual vulnerabilities” for NCII and CSAM.
What the testing could not measure
The report is candid about its main gap. The evaluation “could not directly measure robustness to adversarial modification (e.g. via fine-tuning or low-rank adapters, known as LoRAs).” That is the central risk with open weights. A determined user can retrain a downloaded model to remove its safeguards, and no pre-release test can fully rule that out.
Black Forest Labs argues that its embedded mitigations raise “the expertise, data, and compute barrier” for anyone trying, and it points to Hugging Face and CivitAI improving their moderation of abusive adapters. That is a reasonable argument, but it is an expectation rather than a measured result. The figures also come from the company’s own report, not from an independent publication.
| Stage | Black Forest Labs measure | Partner or standard named |
|---|---|---|
| Pre-training | Filter nude, pornographic and known abuse material from the training set | Internet Watch Foundation |
| Post-training | Targeted fine-tuning against text-to-image and image-to-image attacks | Cinder red-teaming at each checkpoint |
| Deployment | Licences banning misuse; required inference filters; provenance metadata | C2PA; NCMEC reporting on hosted services |
| After release | Misuse monitoring, takedown requests, user bans, safety hotline | Guidance from Thorn, NIST and Ofcom |
| Not covered | Robustness to fine-tuning and LoRA adapters | Stated limitation in the report |
Europe's AI Gap in Numbers
Rombach’s optimism argument rests on a diagnosis: Europe is behind, and fear makes it worse. The first half of that is not in dispute. The official figures show a gap in models, in capital and in valuations, and they explain why Black Forest Labs keeps a second base in California.
The model gap
The Stanford AI Index 2026 says the United States “produced 59 notable models in 2025 to China’s 35.” Model production “remains concentrated in the U.S. and China”, and the index notes that open-source contributions from the rest of the world are “now outpacing Europe” on GitHub.
Europe does have competitive labs. AFP names France’s Mistral alongside Black Forest Labs. Mistral raised a €3 billion Series D this month at a post-money valuation above €21 billion. Both companies, though, are still described as trailing OpenAI, Anthropic and China’s DeepSeek.
The capital gap
The money gap is wider than the model gap. The AI Index puts US private AI investment at $285.9 billion in 2025, “more than 23 times the $12.4 billion invested in China”. It also counts 1,953 newly funded US AI companies in 2025, “more than 10 times the next closest country”.
The European Commission’s own Competitiveness Compass puts it starkly: “the share of global venture capital funds raised in the EU is only 5%, compared to 52% in the US and 40% in China.” It adds that “every year EUR 300 billion of savings from Europeans are invested in markets outside the EU.”
Banks versus venture capital
AFP also spoke to Zach Meyers, director of research at CERRE, a Brussels think tank. He called access to capital the “biggest barrier” to AI growth in Europe. Most European companies rely on bank financing, and banks are more cautious than venture investors about businesses that may take years to turn a profit.
“That is not how technology markets work, particularly with a really nascent technology like AI,” Meyers told AFP. The Commission agrees. Its Compass says the EU “is excessively reliant on bank debt financing” even though its household saving rate was 65% higher than the US rate in 2022. Europeans save plenty. The savings just do not reach risky technology firms at home.
The valuation gap
Valuations show the same pattern at the start-up stage. According to PitchBook’s Q2 2026 European Venture Report, as reported by Tech Funding News, the median pre-money valuation of a European AI startup is €8.3 million, against €64.2 million in the United States.
Dividing one by the other gives a ratio of about 7.7. A European founder at the median raises money on a valuation roughly one-eighth of a US peer’s, so every euro raised costs more equity. This is the practical meaning of Rombach’s “certain ingredients”. It also explains why so many European founders, his included, build strong ties to Silicon Valley investors.
Why Black Forest Labs kept one foot in California
Black Forest Labs is the gap in miniature. It was founded by German researchers, is headquartered in Freiburg, and was funded first by Andreessen Horowitz in Silicon Valley. AFP notes that “the firm’s ties to the United States and its tech scene are also strong”, with a second headquarters in San Francisco.
That is not a contradiction of the optimism argument. It is its evidence. A European deep-tech company that wanted to move quickly in 2024 found its lead investor in California, and it has kept a US base since. Rombach is asking Europe to become the kind of place where the next lab does not need to.
| Measure | Figure | Source |
|---|---|---|
| Notable models released, 2025 | US 59, China 35 | Stanford AI Index 2026 |
| Private AI investment, 2025 | US $285.9bn, China $12.4bn | Stanford AI Index 2026 |
| Share of global VC funds raised | US 52%, China 40%, EU 5% | European Commission, Jan 2025 |
| European savings invested outside the EU | €300bn a year | European Commission, Jan 2025 |
| Median AI startup pre-money valuation | Europe €8.3m, US €64.2m | PitchBook Q2 2026, via Tech Funding News |
| Black Forest Labs valuation | $3.25bn post-money | Series B announcement, Dec 2025 |
From Images to Robots: Black Forest Labs' Physical AI Push
AFP notes that the company “is now also expanding into physical AI”. That expansion changes the safety conversation, because a model that moves a robot arm acts on the world rather than drawing it. Black Forest Labs has published unusually detailed material on the work.
FLUX 3 and the Audi test
On 23 July 2026 Black Forest Labs unveiled FLUX 3, a model trained jointly on images, video and audio. The same day it published FLUX 3 x mimic, describing FLUX-mimic, a video-action model built with the Zurich robotics start-up mimic. The company says FLUX-mimic runs “robots that have been tested and deployed at Audi.”
The target is work that traditional automation has avoided. The blog lists kitting parts into trays, inserting control units into tight fixtures and handling “soft, flexible materials like seals and cables”. Christoph Schneider of the Audi Production Lab is quoted as saying the robots solved “complex soft-body manipulation work that would have been simply impossible with conventional robotics.”
The numbers the company published
The thesis is that a model which can generate realistic video has had to learn how the world behaves. Black Forest Labs says video prediction accounts for “over 95% of the total compute costs” of training FLUX 3. Audio is cheap by comparison, at “less than 0.5% of the tokens in a 720p video with audio”.
Adding action prediction briefly hurt the model. Human ratings of video quality “initially fell by up to 10%”, then recovered fully after 3,500 training steps. For speed, the backbone runs “in less than 80ms on a single NVIDIA RTX 5090 GPU”, which the company compares with human visual reaction time. The full robot system reacts in 101 milliseconds.
FLUX 3 Action goes open
Five days before the AFP interview, Black Forest Labs released FLUX 3 Action, “an open-weight 7B world-action-model”. On the RoboLab-120 simulation benchmark, its single-step checkpoint reaches a 38.3% success rate. NVIDIA’s Cosmos 3 Nano, a 16B model, reaches 36.8%, and Ï€0.5 reaches 28.0%.
A guidance-distilled version raises the score to about 42.2%, though the company notes it is slower. The release follows the same pattern as FLUX in images: publish the weights and the fine-tuning recipe so others can build on them. The report says it is published “along with the weights to make this process transparent.”
| Robot policy (RoboLab-120) | Open weights | Parameters | Success rate |
|---|---|---|---|
| FLUX 3 Action, guidance-distilled | Yes | 7B | 42.2% |
| FLUX 3 Action, single step | Yes | 7B | 38.3% |
| Cosmos 3 Nano policy (NVIDIA) | Yes | 16B | 36.8% |
| π0.5 | Yes | 3.3B | 28.0% |
| GR00T N1.6 | Yes | 3B | 7.2% |
Why physical AI changes the safety question
Image models raise harms of content: deepfakes, abuse imagery and copyright. Robot models raise harms of action. A policy that misreads a scene can damage equipment or injure a worker, and the computer vision that guides it has to cope with conditions a benchmark never shows.
Black Forest Labs’ open approach therefore faces a new test. An open robot policy lets factories, universities and start-ups inspect and adapt it, which supports Rombach’s transparency point. It also means safety depends on how each deployer integrates it, with its own guards, stops and testing. Industrial robotics already has safety standards for that; the novelty is a learned model sitting inside the loop.
Is Open Really Safer? Weighing Black Forest Labs' Claim
Rombach’s line that open weights “increase transparency, increase safety” is the most contested part of the interview. It is worth separating the parts that the evidence supports from the parts that remain a bet.
The case for transparency
The strongest argument is about oversight. When weights are public, independent researchers can test a model directly rather than relying on a vendor’s summary. Black Forest Labs’ comparison against Alibaba, Tencent and ByteDance models was only possible because those models were also open.
Rombach’s other point concerns concentration. At the G7 he said that “a climate of fear around open technology, or a focus on suppression over diffusion, will leave the world reliant on a handful of firms for critical infrastructure.” For European businesses and governments, that is also an argument about strategic independence. Open models can be hosted in Europe, on European infrastructure, under European law.
The case against
The counter-argument is permanence. A hosted model can be patched or withdrawn overnight. An open model cannot, as Black Forest Labs itself admits. If a jailbreak or a harmful fine-tune emerges, copies already downloaded stay in circulation.
The company’s February report measured resistance to prompts, not to retraining. Its most permissive models are also its smallest and most portable, which means they are the easiest to modify on ordinary hardware. The claim that open weights increase safety is therefore strongest for transparency and research, and weakest for misuse by a determined actor.
Different risks from the September incidents
The September incidents that shifted the debate involved something different. OpenAI’s escapes came from agentic language models acting inside research systems, reaching networks they were meant to be kept away from. That is a containment problem at the frontier labs, not a question of open image weights.
Critics who fold the two together overstate the case against Black Forest Labs. Rombach may understate it in turn. FLUX 3 Action moves the company into models that act, and “world models” are exactly where agent-style risks begin to apply. The debate over how an AI slowdown could be enforced is mostly about compute and language models today, but robotics will not stay outside it.
Our reading
Rombach is right that fear alone is a poor industrial strategy, and his company has done more than most to show its safety work. The numbers are self-reported, however, and the untested risk, retraining to strip safeguards, is the one that matters most for open models.
A fair verdict is that Black Forest Labs has made a credible case that openness and safety can coexist for image generation, with measured safeguards. It has not shown that openness makes models safer in every respect, and nobody yet has. Optimism is a reasonable position; certainty is not.
| Question | Open-weight model | Hosted closed model |
|---|---|---|
| Can outsiders test it directly? | Yes, anyone with the weights | Only through the vendor’s interface |
| Can a flaw be fixed everywhere? | No; old copies keep circulating | Yes, by updating the service |
| Can safeguards be removed? | Yes, with fine-tuning skill and compute | Only by jailbreaking the live service |
| Can it run on your own servers? | Yes, subject to licence | Rarely |
| Who carries deployment controls? | Largely the deployer | Largely the vendor |
The EU AI Act and Black Forest Labs
Rombach did not attack the EU AI Act in the AFP interview, and his company’s documents suggest it is working within it. Regulation is often blamed for Europe’s position, but the 2026 changes to the Act are more mixed than that story allows, and one of them lines up with Black Forest Labs’ own safety priorities.
What the Omnibus changed
On 7 May 2026 the Parliament and Council reached a provisional deal on the Digital Omnibus on AI, a package of amendments to the Act. The Council gave its final approval on 29 June. As summarised by Gibson Dunn, high-risk obligations for stand-alone Annex III systems move to 2 December 2027, and those for AI embedded in regulated products move to 2 August 2028.
That delay is a concession to the competitiveness argument that Rombach and others make. Europe has decided that some of its rules were landing before firms could comply. It did not, however, loosen everything.
A new ban on abusive imagery
The Omnibus adds a prohibition on AI systems that generate or manipulate “non-consensual intimate images, video, audio or similar material, or child sexual abuse material”. It applies where that output is “a reasonably foreseeable and reproducible outcome, without requiring significant technical modification”, with a transition period until 2 December 2026.
This matters directly for Black Forest Labs. The phrase “without requiring significant technical modification” maps onto the company’s approach, where safeguards are built into the released model and the untested risk is retraining. At the G7, Rombach said the company welcomed the US, EU and UK developing “new legal strategies to combat sexual deepfakes”. The EU has now done so.
Black Forest Labs’ own disclosures
The company runs a Transparency Hub, last updated on 1 August 2026. It hosts model documentation and training data summaries for its general-purpose AI models under the EU AI Act, and a training data disclosure under California’s AB 2013.
Its legal pages include separate EU developer terms, EU API service terms and EU self-hosted commercial licence terms, all updated on 26 August 2026. The company’s Responsible AI Development Policy sits alongside them. A lab arguing for optimism is, in practice, doing the compliance paperwork European law requires.
Labelling from August 2026
Article 50 transparency duties have applied since 2 August 2026, including machine-readable marking of AI-generated content. Systems already on the market before that date have until 2 December 2026 for the marking. We looked at whether AI content labels will help or hinder deepfake detection.
Black Forest Labs already applies C2PA provenance metadata on its hosted services and points developers to C2PA in its open-weight repositories. For businesses using FLUX outputs in the EU, that makes the marking duty easier to meet, but only if the metadata survives each tool in the chain.
| Date | EU AI Act milestone | Relevance to image and video models |
|---|---|---|
| 2 Aug 2025 | General-purpose AI model obligations apply | Documentation and training data summaries |
| 7 May 2026 | Provisional Digital Omnibus deal | Adds the abusive imagery ban; delays high-risk rules |
| 29 Jun 2026 | Council’s final approval | Omnibus text settled |
| 2 Aug 2026 | Article 50 transparency duties apply | Machine-readable marking of AI output |
| 2 Dec 2026 | Marking grace period and ban transition end | Existing systems must comply |
| 2 Dec 2027 | Annex III high-risk obligations | Mostly outside creative image generation |
| 2 Aug 2028 | High-risk rules for AI in regulated products | Relevant to robot and machinery use |
The Scorsese Deal and Hollywood's Pushback
AFP mentions that Black Forest Labs “recently hit the headlines” when it named Martin Scorsese as an adviser, “prompting an angry response from some in Hollywood”. That row shows a different kind of fear from the one in the safety debate, and it is harder for optimism alone to answer.
What was announced
On 2 June 2026 the company published a page, Martin Scorsese × Black Forest Labs, with a video of the director storyboarding with FLUX. According to Tech Times, the scenes came from his next film, What Happens at Night, starring Leonardo DiCaprio and Jennifer Lawrence.
“Now, with this tool, I can share what I’m visualizing more clearly and efficiently to my creative team—the production designer, art designer, and cinematographer,” Scorsese says on the page. The link runs through investors. BroadLight Capital, co-founded by Scorsese’s manager Rick Yorn, is a Black Forest Labs investor, and CAA co-founder Michael Ovitz, an early angel investor and adviser, helped broker the partnership.
The Art Directors Guild responds
On 9 June the Art Directors Guild (IATSE Local 800) issued a statement, reported by ComingSoon and others. It said Scorsese was “turning his back on the human artists who throughout his career have helped him create his most memorable works.”
The guild’s objection went to how the tools are built. It argued that generative AI can only produce “cinematic intelligence” by “ingesting large swaths of copyrighted work, likely scraped from the internet without consent, credit, compensation, or transparency”, and called the idea that it could outshine human artists “a betrayal of the collaborative nature of cinema.”
Artists’ objections
Individual artists were sharper. Concept artist Karla Ortiz said Scorsese “throws every single storyboard artist he’s ever worked with under the bus”. Animation director Sam Deats said: “It takes literally seconds for me to storyboard a shot, there is absolutely no reason to need AI built on the stolen work of millions of artists.”
Whether Scorsese has a personal financial stake is unclear. Tech Times reported that Black Forest Labs “did not confirm or deny it”. The row echoes other recent disputes over creative work, such as venues banning AI generated artwork.
What it means for the optimism argument
Rombach’s optimism is aimed at policymakers worried about catastrophic risk and at investors worried about Europe’s position. The Hollywood reaction is about something more immediate: jobs, consent and payment for training data. The phrase “risk and fear” covers both, but they need different answers.
Transparency about safety testing does not settle questions about training data or livelihoods. Black Forest Labs publishes training data summaries under EU and California law, which is a start. Winning over creative workers will likely need licensing and compensation arrangements, not just a more upbeat mood.
What Black Forest Labs' Stance Means for Businesses
For UK and European businesses, the interview is more than a policy argument. Black Forest Labs sells models that companies use every day, and its stance on openness shapes how they can be deployed. Here are the practical points we would raise with any client considering FLUX or a similar model.
Check the licence before you build
The FLUX family mixes licences. The Apache 2.0 models can be used commercially without a separate agreement. FLUX.1 [dev], FLUX.2 [dev] and the 9B [klein] models are non-commercial by default, so commercial self-hosting needs Black Forest Labs’ paid licence or API access.
Getting this wrong is a common and avoidable mistake. Record which model and licence each workflow uses, and check again when you upgrade, because a newer version can carry different terms. If your team is weighing hosted against self-hosted AI, an AI strategy review should cover licensing alongside cost and data location.
Keep the safeguards switched on
Black Forest Labs’ own report says deployment-time moderation removes most residual risk, and its licences require inference filters. Self-hosting teams sometimes disable filters for speed or convenience. Do not. Keep them on, log prompts and outputs where the law allows, and run your own tests against misuse before launch.
Our guide to AI red-teaming before launch sets out what to test. For image tools, that includes attempts to create images of real people, to remove clothing and to generate brand or trademark misuse.
Plan around EU dates, even from the UK
EU AI Act duties follow the output, not only the provider’s address. A UK firm whose AI-generated images reach EU customers should plan for Article 50 marking and keep C2PA metadata intact through its tools. The 2 December 2026 dates for the marking grace period and the abusive imagery ban are close.
If you are unsure whether a use case falls within scope, document your reasoning now. A short register of AI tools, their outputs and their audiences takes an afternoon and makes later compliance work much simpler.
Keep the fear proportionate
Rombach’s broader point applies inside companies too. A risk register that treats every AI tool as an existential threat stops useful work, while one that ignores misuse invites incidents. Rate each use case on its real exposure: who sees the output, what it could harm, and how you would detect a problem.
Most businesses using image generation for marketing, product visuals or internal design face manageable risks with clear controls. The safety fears around frontier agents are real, but they are a different category from a well-configured image tool. Treating them the same helps no one.
Black Forest Labs FAQ
What did Black Forest Labs’ CEO say about Europe?
Robin Rombach told AFP that “the mindset in Europe needs to shift to one of optimism and to one of opportunity, and not to one of risk and fear.” He argued that Europe lacked a startup ecosystem for frontier deep tech and that success needs speed, capital and computing power.
Who founded Black Forest Labs?
Robin Rombach, Andreas Blattmann and Patrick Esser, researchers behind latent diffusion and Stable Diffusion. The company launched on 1 August 2024 in Freiburg, Germany, and also has a headquarters in San Francisco.
Are Black Forest Labs’ FLUX models open source?
Some are open weight rather than fully open source. FLUX.1 [schnell] and the FLUX.2 [klein] 4B models use Apache 2.0. FLUX.1 [dev], FLUX.2 [dev] and the 9B [klein] models use a non-commercial licence, and commercial use needs a paid licence.
Is Black Forest Labs working with Audi?
Yes, through FLUX-mimic, a robot model built with mimic robotics on the FLUX 3 backbone. Black Forest Labs says it runs robots tested and deployed at Audi for tasks such as handling seals and cables.
How much is Black Forest Labs worth?
Its December 2025 Series B raised $300 million at a $3.25 billion post-money valuation. AFP describes the company as valued at around $3.25 billion.
References and Further Reading
Yahoo Tech (AFP): AI startup urges optimism from Europe despite safety fears
Tech Xplore (AFP): AI startup urges optimism from Europe despite safety fears
Black Forest Labs: Our co-founder and CEO urges G7 leaders to back open innovation
Black Forest Labs: Capable, Open, and Safe: Combating AI Misuse
Black Forest Labs: FLUX 3 x mimic
Black Forest Labs: FLUX 3 Action
Black Forest Labs: Our $300M Series B
Black Forest Labs: Announcing Black Forest Labs
Black Forest Labs: Transparency Hub
arXiv: High-Resolution Image Synthesis with Latent Diffusion Models
Stanford HAI: The 2026 AI Index Report
European Commission: A Competitiveness Compass for the EU
Council of the EU: Council gives final green light to simplify AI rules
Gibson Dunn: EU AI Act Omnibus Agreement
Tech Funding News: European AI startups valued at one-eighth of US peers
Tech Times: Martin Scorsese joins Black Forest Labs as AI adviser
ComingSoon: Art Directors Guild criticises Scorsese’s AI support
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
