SDXL
Lead author · 2023
Dual text encoders, multi-aspect training, two-stage refinement. ICLR '24 Spotlight, 4,000+ citations. One of the most widely adopted open-source generation models.
Builds generative models that actually ship.
Co-founder at Black Forest Labs. Lead author of SDXL. Creator of Kontext. Core contributor to the FLUX model family. Self-taught, 12+ production models, and a pattern of finding the highest-leverage research direction in the room and executing it into something real.
About
Generative AI researcher and engineer who consistently identifies high-leverage research directions and pushes them from concept through to shipped production systems.
Co-founded Black Forest Labs, which raised a $300M Series B at a $3.25B valuation. Lead-authored SDXL, one of the most widely adopted open-source image generation models. Created Kontext, designing in-context image editing with a two-person core R&D team over roughly two months. Core contributor across the FLUX.1 and FLUX.2 model lines.
His work spans diffusion and flow-matching systems, transformer architectures (MM-DiT, DiT), large-scale distributed training, data curation, representation learning, and multi-modal model design. The common thread is leverage — finding the few design decisions that change the trajectory of a model line, then executing them quickly.
Came into machine learning without a formal degree, teaching himself deep learning and generative modeling through direct implementation and open-source work. Before that, seven years across visual effects, media production, and software development — which partly explains the strong instinct for both aesthetic quality and practical execution speed.
Work
12 shipped models
CosXL
Creator · 2023
Cosine-scheduled noise scaling and v-prediction for sharper color accuracy. A small, precise intervention with outsized visual impact.
Control-LoRAs
Creator · 2023
Depth, canny, and pose conditioning as lightweight LoRA adapters. Made fine-grained control practical and accessible.
Stable Video Diffusion
Contributor · 2023
Extended latent diffusion into the temporal domain for video generation.
SD3
Contributor · 2024
Next-generation architecture work around rectified flow transformers. ICML '24 Best Paper. Directly informed the Flux MM-DiT design.
FLUX.1
Core contributor · 2024
12B-parameter MM-DiT built from scratch. Rectified flow matching, 3D RoPE, guidance distillation. Shipped dev, schnell, and pro in four months.
FLUX.1.1
Contributor · 2024
Continued iteration on Flux quality, behavior, and deployment characteristics.
Flux Tools
Contributor · 2025
Ecosystem-level product work around the Flux model family — tooling, not just flagships.
Flux Ultra
Creator · 2025
Ultra-high-quality production model. The premium end of the Flux line.
Kontext
Project lead · 2025
In-context image editing via sequence concatenation on a 12B Flux transformer. Core R&D: 2 people, ~2 months. Introduced KontextBench.
FLUX.2
Core contributor · 2025
Native multi-reference and editing in pretraining. Massively scaled paired data. Extended the Kontext line into the model stack itself.
Klein
Contributor · 2025
Efficient small-scale inference model. Rounds out the family from the deployment end.
Expertise
Technical focus
Diffusion & flow matchingMM-DiT / DiT architecturesRectified flowRepresentation learningLarge-scale GPU trainingData curation & filteringIn-context learningMulti-modal architecturesLoRA & adapter methodsProgressive curriculum design
Models shipped
12+
to production
Lead roles
SDXL · Kontext
Citations
4,000+
SDXL alone
Path
Self-taught
no formal degree
Experience
Co-Founder & Research Scientist — Black Forest Labs
2024 — Present
San Francisco / Freiburg · $300M Series B at $3.25B valuation
- Multi-modal research: representation learning and omni-model architectures across image, video, audio, and text.
- Project lead on Kontext — in-context image editing via sequence concatenation on a 12B Flux transformer. Core R&D team of 2 people, ~2 months. Introduced KontextBench (1,026 pairs, 5 task categories).
- Core contributor to FLUX.1: a 12B MM-DiT with rectified flow matching, 3D RoPE, and guidance distillation, shipped across dev/schnell/pro in four months.
- Core contributor to FLUX.2: native multi-reference and editing capabilities in pretraining, massively scaled paired training data.
- Creator of Flux Ultra (ultra-high-quality generation) and contributor to Klein (efficient small-scale inference).
- Recent published work: Self-Flow, a self-supervised flow matching framework with 2.8× faster convergence.
Applied Model Lead — Stability AI
2022 — 2024
Los Angeles / Remote · Promoted from ML Engineer to Lead in 8 months
- Lead author of SDXL — dual text encoders (CLIP ViT-L + OpenCLIP ViT-bigG), multi-aspect training with micro-conditioning, two-stage refinement pipeline. ICLR 2024 Spotlight. 4,000+ citations.
- Created CosXL: cosine-scheduled noise scaling and v-prediction for improved color accuracy and quality.
- Created Control-LoRAs: depth, canny, and pose conditioning as lightweight adapters.
- Contributed to Stable Video Diffusion and SD3 / Scaling Rectified Flow Transformers (ICML 2024 Best Paper).
Self-Directed ML Research
2021 — 2022
Los Angeles, CA
- Taught himself deep learning and generative modeling from scratch — diffusion models, UNet architectures, training methodology — through hands-on implementation, tinkering, and open-source mentorship. No formal degree or coursework.
Technical Roles — VFX, Media Production, Software
2015 — 2021
Los Angeles, CA
- Seven years of technical problem-solving across complex production environments. Built the rapid-learning instincts and quality intuition that later translated unusually well to ML research.
Publications
SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis
ICLR 2024 Spotlight · Lead author
Scaling Rectified Flow Transformers for High-Resolution Image Synthesis
ICML 2024 Best Paper · Contributor
FLUX.1 Kontext
2025 · Project lead
Self-Flow: Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis
arXiv 2603.06507 · Author
Trajectory
2015 — 2021
Seven years across VFX, media production, and software. Built technical instincts and rapid-learning habits in complex production environments.
2021 — 2022
Pivoted into self-directed ML research. Taught himself the full stack from scratch without a degree, driven by conviction that generative models were the frontier worth investing in.
2022 — 2024
Joined Stability AI, quickly became lead author of SDXL, and was promoted to Applied Model Lead within eight months. Shipped multiple production models across image and video.
2024 — Now
Co-founded Black Forest Labs. Led Kontext, contributed to the FLUX model family, and continues pushing multi-modal research at the frontier.