Member of Technical Staff – Safety
Stellenbeschreibung
About Black Forest Labs
We’re the team behind Latent Diffusion, Stable Diffusion, and FLUX—foundational technologies that changed how the world creates images and video. We’re creating the generative models that power how people make images and video—tools used by millions of creators, developers, and businesses worldwide. Our FLUX models are among the most advanced in the world, and we’re just getting started.
Headquartered in Freiburg, Germany with a growing presence in San Francisco, we’re scaling fast while staying true to what makes us different: research excellence, open science, and building technology that expands human creativity.
Why this role
FLUX models ship as open weights, which gives anyone full access to the model. Our safety efforts focus on suppressing harmful behaviors and mitigating the risk of misuse, so safety has to be built into the models themselves, not bolted on afterwards.
As a Member of Technical Staff on our safety research team, you’ll research and ship methods that make FLUX models safe by design: from what goes into the training data to how the models are mid- and post-trained, across images, video and the text that conditions them. This is a role with real ownership: you won’t just do research on model safety, you’ll shape it. The methods you build ship in FLUX models with over 75 million downloads on Hugging Face and in production at some of the largest platforms in the world.
What you’ll work on
-
Develop mid- and post-training safety methods for FLUX models, such as safety SFT, preference optimization and RL with safety rewards, and concept erasure.
-
Build safety data filtering that runs at pretraining scale: classifiers and pipelines that screen billions of images, video clips and captions, with measured Pareto frontiers and precision/recall trade-offs.
-
Make safety hold up for open-weight releases, including robustness to adversarial prompts and to downstream fine-tuning.
-
Extend safety to the text side of the stack: prompts, captions and the text encoders that condition our models.
-
Design evaluations and red-teaming that measure harmful generations alongside image quality and prompt following, so safety gains don’t come at the cost of performance.
-
Work directly with pretraining and post-training teams so your methods land in the release model, and publish where it makes sense.
What we’re looking for
-
Hands-on research experience in visual safety: safety of text-to-image, text-to-video or vision-language models, such as harm classifiers, safe generation methods or safety benchmarks.
-
Proven experience in at least one of:
-
Large-scale safety data filtering: building and running classifiers or filtering pipelines over hundreds of millions of samples or more.
-
Safety mid- or post-training: safety SFT, DPO or RL with safety objectives, or unlearning applied to large generative models.
-
-
Deep familiarity with concept erasure methods (e.g. ESD, UCE, adversarial unlearning) and their failure modes, such as concept recovery and collateral damage to unrelated concepts.
-
Track record of shipping: top-venue publications (NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV) paired with production impact, or unambiguous production wins at a frontier lab.
-
Experience with distributed training at multi-node scale.
-
Strong PyTorch fluency; you write research code that others can build on. You work fluently with coding agents to move faster, and have the depth to verify their output and own the result.
-
Experience running multi-agent workflows: you know which tasks to hand off and which to keep yourself, and you use the speed-up to raise the quality bar, not lower it.
We’d be especially excited if you:
-
Know how to keep safety intact after release, when others fine-tune open weights.
-
Have tackled safety for video generation, where harm can emerge across frames rather than in any single frame.
-
Bring experience with audio safety, especially for jointly generated audio and video, where each modality can amplify harm in the other.
-
Have unlearned concepts from text encoders or built safety-tuned encoders (e.g. Safe-CLIP-style approaches).
-
Have worked on text safety: LLM safety tuning, prompt and jailbreak defenses, or text safety classifiers, ideally across languages.
-
Have applied latent-space safety methods, such as activation steering or safe guidance.
-
Have built VLM-based safety judges or guard models for filtering or evaluation.
How we work together
We’re a distributed team with real offices that people actually use. You’ll join us in Freiburg at least 2 days a week (or one full week every other week). We’ll cover reasonable travel costs to make this possible. We think in-person time matters, and we’ve structured things to make it accessible to all. We’ll discuss what this will look like for the role during our interview process.
Everything we do is grounded in four values:
-
Obsessed. We are a frontier research lab. The science has to be right, the understanding deep, the product beautiful.
-
Low Ego. The work speaks. The best idea wins, no matter who said it. Credit is shared. Nobody is above any task.
-
Bold. We take the ambitious bet. We ship, we do not wait for conditions to be perfect.
-
Kind. People over politics. We treat each other with genuine warmth. Agency without empathy creates chaos.
If this sounds like work you’d enjoy, we’d love to hear from you.
Compensation
EUR 160K-220K + equity, depending on profile
Ähnliche Jobs