The Opportunity
We are solving AI-driven creativity across multiple axes. On the visual generation side, we are building a proprietary, internal diffusion foundation model targeting closed-SOTA performance (rivaling Midjourney, Flux, and Ideogram).
This is not an open-source wrapper, a basic LoRA project, or a set of disjointed research tasks. You will work at the scale and technical depth of a foundational closed diffusion lab - owning end-to-end post-training, intelligence, visual taste, and custom controllability short of raw $20M scratch pre-training runs.
We are opening two specialized research seats for exceptional machine learning researchers and engineers to take complete ownership of core pillars of our visual stack:
🟢 Seat 1: Applied ML Researcher - RL, Preference & Reward Modelling
Core Mandate: Take full ownership of the preference, evaluation, and reward alignment side of our model stack.
What You’ll Own: Diffusion RL: Apply online RL, DPO, GRPO, and successor algorithms to real-world visual quality, prompt adherence, and controllability.
Taste & Reward Modelling: Build and train custom reward models for aesthetic scoring, artifact elimination, instruction-following, and human preference.
Evaluation Suites: Design evaluation harnesses and datasets that directly correlate with human visual taste.
Controllability & Adapters: Develop custom hypernetworks, ControlNet/IP-Adapter extensions, and fine-tuning pipelines for precise visual execution.
🔵 Seat 2: Applied ML Researcher - Model Distillation & Inference Efficiency
Core Mandate: Take full ownership of model speed, distillation, and compute optimization - making our models generate significantly faster and cheaper without sacrificing visual quality.
What You’ll Own: Few-Step Distillation: Lead few-step and progressive distillation on our internal foundation models and baseline architectures.
Inference Optimization: Implement low-precision training (FP8), quantization, torch.compile paths, and custom CUDA/Triton kernels.
Distributed Systems: Scale and maintain distributed training infrastructure on multi-GPU clusters as the core centre of gravity for our efficiency stack.
What We Are Looking For
Hands-on Diffusion Expertise: Demonstrated track record with hands-on SFT, LoRA, and fine-tuning of diffusion models or DiTs.
Distributed Training: Direct experience training large models on multi-GPU/multi-node clusters using PyTorch, FSDP, DDP, or DeepSpeed.
Track Record & Credentials: Final-year Ph.D. (or equivalent industry experience) specifically in image generation/diffusion. A first-author publication in top computer vision / ML venues (CVPR, ICCV, ICML, ICLR, or NeurIPS) or time at a leading diffusion lab (Adobe, Runway, Ideogram, Snap, etc.) is strongly valued.
High Intensity: Early-stage team pace with direct, high-impact research ownership