FLUX models are released as open weights, so anyone can access the full model. Because of that, safety has to be designed into the models from the start rather than added later.
This is a Member of Technical Staff role on the safety research team. You will research and ship methods that make FLUX models safe by design, covering training data, mid- and post-training, and the images, video and text that condition them. You will have real ownership over the direction of model safety.
Responsibilities
Develop mid- and post-training safety methods for FLUX, such as safety SFT, preference optimization, RL with safety rewards and concept erasure.
Build data filtering pipelines and classifiers that run at pretraining scale over billions of images, video clips and captions, and measure their Pareto frontiers and precision/recall trade-offs.
Keep safety effective in open-weight releases, including against adversarial prompts and downstream fine-tuning.
Extend safety work to text: prompts, captions and the text encoders that condition the models.
Design evaluations and red-teaming that track harmful generations alongside image quality and prompt following, so safety gains do not reduce performance.
Work directly with pretraining and post-training teams to get your methods into the release model, and publish where it makes sense.
Requirements
Hands-on research experience in visual safety for 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; or safety mid- or post-training, including safety SFT, DPO, RL with safety objectives or unlearning on large generative models.
Deep familiarity with concept erasure methods such as ESD, UCE and adversarial unlearning, including their failure modes like concept recovery and collateral damage to unrelated concepts.
A record of shipping: top-venue publications (NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV) combined with production impact, or clear production wins at a frontier lab.
Experience with distributed training at multi-node scale.
Strong PyTorch skills, and you write research code others can build on.
You work fluently with coding agents to move faster, and you can check their output and take ownership of the result.
Experience running multi-agent workflows: you know which tasks to delegate and which to keep, and you use the speed gain to raise quality.
Especially welcome if you: know how to keep safety intact after release when others fine-tune open weights; have worked on video safety, where harm can appear across frames rather than in one frame; have worked on audio safety, especially for jointly generated audio and video; have unlearned concepts from text encoders or built safety-tuned encoders such as Safe-CLIP-style approaches; have worked on text safety, such as 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; or have built VLM-based safety judges or guard models for filtering or evaluation.
Conditions
The team is distributed but has offices that people use. You will be in Freiburg at least 2 days a week, or one full week every other week. The company covers reasonable travel costs for this. Specifics for the role will be discussed during interviews.
The team works by four values: obsessed with getting the science and product right, low ego with shared credit and the best idea winning, bold in shipping without waiting for perfect conditions, and kind with genuine warmth and empathy.
Pay is EUR 160K-220K plus equity, depending on profile.
Equal opportunity applies to all candidates.
Benefits
Equity, in addition to the EUR 160K-220K base pay range.