Adewale

Generative AI Workflow Engineer

Lumen Creative • Freelance

May 2022 - Present • 4 yrs 6 mos

I design and run image generation pipelines for creative studios, fashion labels and social media agencies that need AI characters and product visuals at scale. My day to day work is building ComfyUI workflows from scratch, training LoRA models on small client photo sets, and keeping the setups stable as models and custom nodes change. A typical project starts with a dataset review, where I check that the training photos cover different angles, lighting and expressions. Then I train and test several versions of the LoRA until the face holds across close ups, full body shots and different outfits. Next I build the generation workflow with face detail, upscaling and batch output, and I hand over the exported file with a written guide so the team can run it alone. I also repair broken workflows after updates, fixing missing custom nodes, wrong model paths and out of memory errors by changing settings, tiling the heavy steps or moving jobs to rented cloud GPUs. Other duties include comparing Stable Diffusion XL and Flux for each brief, turning finished stills into short image to video clips, and documenting every setup so new team members can learn it quickly. I also advise clients on which hardware to rent, how many images to generate per batch and how to name and store their files, so nothing gets lost between projects. Every delivery includes a short test run recorded on screen, so the client can see the workflow working before I close the job.

Junior Machine Learning Engineer

Ateliers • Freelance

Jan 2021 - Jul 2022 • 1 yr 6 mos

I worked on the computer vision team, building and testing models that sort and tag product photos for online retailers. My main tasks were cleaning and labelling image datasets, writing Python scripts to resize, crop and balance training data, and training image classifiers in PyTorch. I ran experiments on a shared GPU server, tracked the results in simple reports and helped the senior engineers decide which model version to keep. When a model failed on dark or blurry photos, I found the weak spots in the data, added better examples and retrained it until the errors dropped. I also packaged finished models with Docker so the team could run them on any machine, and I wrote step by step setup notes that new colleagues used on their first day. Part of the job was fixing broken training runs: reading error logs, solving out of memory crashes, and correcting wrong library versions. In my last months I started testing early diffusion models for generating extra training images, which is where my interest in AI image generation began. I presented my results to the team every two weeks and kept all code in Git with clear commit messages. I took part in code reviews, learned how to read other people’s pull requests and how to explain technical problems in plain language to non technical colleagues. I also helped plan the data collection for new product categories, estimating how many photos each class needed and how long labelling would take, which kept the project on schedule.