PhD Research Scientist Intern - Reinforcement Learning for Diffusion Modelling

Join the team redefining how the world experiences design. Our flagship campus is in Sydney, Australia but Austria is home to part of our European operations. You have choice in where and how you work, we trust our Canvanauts to choose the balance that empowers them and their team to achieve their goals.

We’re looking for current PhD students ready to bring their research into the real world and help shape the culture of AI at Canva. Our full-time, 16 week AI Research Internship starts in September. During your internship, you’ll work directly with Canva’s AI team on a live, industry-scale project, turning part of your PhD journey into real world impact. You’ll gain hands-on experience with real data, production infrastructure and real deadlines, while learning from and working alongside the researchers and engineers creating Canva’s next generation of AI-powered experiences.

What you'd be doing in this role:

• Designing and validating a rubric-guided, per-layer VLM judge for RGBA layer decomposition, calibrated against human evaluations.
• Building VLM-based methods for automatic, human-aligned evaluation of multi-layer designs.
• Turning VLM-based evaluators into reward functions to train generative models in a reinforcement learning setting.
• Distilling those judges into lightweight reward models that score layered images from learned representations, at a fraction of the inference cost.
• Collaborating with research, engineering, and product teams to move findings toward production and Canva's layered-generation roadmap.
• Contributing to the broader research community through publication where results support it.

You're probably a match if:

• You're currently completing a PhD, ideally third year or later.
• A strong diffusion or flow-matching background, with hands-on policy-gradient RL for generative models (GRPO, PPO, DPO or similar).
• Experience fine-tuning VLMs (e.g. with LoRA) and designing prompts or rubrics for evaluation tasks.
• Reward modelling experience, preference optimisation, pseudo-labelling, distillation.
• You can read a recent paper and reproduce it quickly.
• You communicate technical work clearly, in writing and in presentations.
• You enjoy working closely with researchers and engineers on hard problems.
• Juggle several threads at once, drop into a new one without losing the last.
• Set your own priorities on a daily basis and between checkpoints.

Nice to have:

• PyTorch at scale, and the ability to write research code for data processing, training and evaluation.
• Multi-GPU training (FSDP, DeepSpeed) and evaluation-harness engineering.
• Layered or RGBA generation, matting, or inpainting experience.
• Familiarity with reward-hacking and score-compression diagnostics, or human-evaluation design.
• Publications or open-source contributions in generative modelling, RLHF, or multimodal models.

What you should aim to take away:

• Publishable and patentable contributions based on the work you’ve done.
• A paper draft covering said work, with support on publication strategy.
• Compute, base checkpoints, preference data and annotation budget, provided.
• Four mentors: a coach for weekly 1:1s, plus a specialist lead on each workstream.
• Work that feeds directly into a product used by hundreds of millions of people.

We make hiring decisions based on your experience, skills and passion, as well as how you can enhance Canva and our culture. When you apply, please tell us the pronouns you use and any reasonable adjustments you may need during the interview process. We celebrate all types of skills and backgrounds at Canva so even if you don’t feel like your skills quite match what’s listed above - we still want to hear from you! Please note that interviews are conducted virtually.