Applied Research Engineer, Evaluation

Black Forest Labs; Zürich; Hybrid; CHF 139,000 - 203,000; Full time

Job summary

Black Forest Labs in Freiburg or Zurich seeks a Senior Evaluation Engineer to build and maintain the evaluation stack for FLUX models, including automated benchmarking and LLM-as-judge pipelines. You will lead human panels and refine metrics for multimodal evaluation across image, video, and beyond.

The role requires deep expertise in Python, PyTorch, and statistics, plus experience with benchmarking and ML infrastructure. Collaboration with RLHF researchers is essential to improve checkpoints.

Qualifications

• Advanced degree (MS/PhD) in a relevant field or equivalent practical experience.
• Practical experience evaluating generative models with hands-on work.
• Strong Python and software engineering skills for reliable pipelines.
• Working knowledge of PyTorch and ML infrastructure practices.
• Solid grasp of statistics, sample sizes, and inter-annotator agreement.
• Experience building or operating LLM-as-judge systems.
• Familiarity with benchmarking datasets and standard benchmarks.

Responsibilities

• Build and maintain the evaluation stack that every model checkpoint passes through.
• Automate benchmarking and LLM-as-judge pipelines and manage human panels.
• Design broad evaluation across taxonomy: prompt adherence, aesthetics, safety.
• Ensure eval results remain stable over time and improve them when needed.
• Benchmark FLUX against frontier models and interpret differences meaningfully.
• Identify outputs where quality falls short not captured by aggregates.
• Collaborate with RL and finetuning researchers to set target metrics.

Skills

Python

Research evaluation

Data analysis

ML pipelines

Education

MS/PhD in a relevant field

Tools

PyTorch

Prolific

Job description

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 videos. 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

Frontier image and video models are now good enough that the hard question is no longer whether they work, but whether a new one is actually better, and how you'd know. Aesthetics, prompt following, editing fidelity, text rendering, character consistency: none of these collapse into a single number, and public benchmarks have fallen well behind the models they're meant to measure.

You'll work on the system BFL uses to answer that question. Every FLUX checkpoint, every finetuning and RL run, and every competitor release goes through our evaluation stack, and you'll help build and sharpen it. Product Research relies on what it shows to decide what to fix next, and to confirm the fix worked.

It gets harder from here. FLUX 3 is our first multimodal model, spanning image, video, audio and action, and each new modality needs its own way of being measured.

What You'll Work On

What you focus on will depend on where you're strongest and what the team needs most when you join.

In scope:

• Building and maintaining the evaluation stack that every model checkpoint has to pass through and prove itself on, from automated benchmarking and LLM-as-judge pipelines, to human evaluation through panels (Prolific and otherwise).

• Designing the breadth of evaluation across the taxonomy, from prompt adherence, editing fidelity, aesthetics, text rendering, subject and character consistency, to safety.

• Evaluating the evaluation stack itself: checking that human and automated evals agree, that results stay stable over time, and improving them when they don't.

• Benchmarking FLUX against the frontier models continuously, while obtaining a meaningful perspective on where we are better or worse than competitors and what such differences really mean.

• Finding the ways in which the quality of the model's output falls short that are not captured by aggregate numbers.

• Closing the loop, ideally, with the RL and finetuning researchers, by defining target metrics for them to aim for, and verifying that a new checkpoint improved the weakness without introducing new ones.

• Designing how best to exploit the signals we already capture through the API and MCP, and what new signals we should capture, with help from research, data engineering and legal.

• Maintaining the usefulness of existing metrics as their natural lifecycle brings up concerns of gaming and overfitting, and anticipating this for new metrics.

What We're Looking For

You have made meaningful contributions to the evaluation of generative models already. You understand that a benchmark figure without careful consideration of sample size and judge agreement with humans is at best uninformed optimism and at worst dangerously misleading, and you have been the person that found that inconvenient regression that nobody wanted to talk about.

You likely have:

• Advanced degree (MS/PhD) in a relevant field or demonstrated equivalent practical experience

• Practical experience with evaluating generative models, in a hands‑on way. You may have worked with diffusion or flow‑matching based image or video models already, but the capability to evaluate them in earnest transfers to other modalities too

• Very strong Python and software engineering skills, with pipelines that run reliably and that other people can read if needed

• Working knowledge of PyTorch and modern ML infrastructure and practices; you are not a pure Python developer, you know what the modern ML code looks like

• Strong grasp of statistics and its shortcomings in practice, and how to work with those; sample sizes, inter‑annotator agreement, confidence levels, etc., and a healthy skepticism about what those numbers say about the world

• Practical experience with building and operating LLM‑as‑judge systems, and understanding their limitations

• Working knowledge of modern benchmarking practices and of standard datasets and benchmarks in the field already; you have an idea of which to trust and which to disregard

• The inclination to look at the actual outputs your system is judging, and to care about what is there beyond an aggregate score

We'd be especially excited if you:

• Have practical experience with human evaluation, including study design and execution via panels, and analysis of disagreements and annotator performance

• Have worked where evaluation meets training (reward models, preference data, RLHF, DPO or similar) and know that whatever you optimize for is what the model learns

• Have evaluated agentic or multi‑step systems and know how hard it is to attribute success or failure to a single step

• Have built tools that people actually want to use internally

• Have made public contributions to the field: papers, open source, or otherwise

How We Work Together

We're a distributed team with real offices that people actually use. This role is based in Freiburg or Zurich, where you'll join us at least 3 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 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 : €150K-€220K base + equity depending on experience