The unpredictability of machine learning proved to be both challenging and promising in my commercial visual work.

Upgrading CGI Quality

To meet client demands for high-quality visuals, I learned that in-painting straight onto high resolution Blender renders, give me the most control.

Before After

Masking in Latent Space

I also learned to use masking in the latent space to introduce elements that were not recognised by the ML model.

An effective strategy for rendering these ‘unlearnable’ elements into photographically-enhanced CGI environments, whilst maintaining their unique design features.

Before After

Overcoming Prompt Engineering Limitations

While the surprising and creative outcomes from text-to-image models can be beneficial, I found pure prompt engineering in text-to-image models restrictive. Leading me more towards image-to-image workflows to cater more effectively to my clients expectations.

Blender Render
IMG2IMG SDXL Render *with controlled variation
Blender Render
IMG2IMG SDXL Render *with controlled variation

Enhancing Workflow Efficiency

There are plenty ways to boost the entire workflow efficiency for stock photography using tools like ComfyUI, SDXL, ControlNet Depth, and IPAdapter, facilitating a quicker turnaround for projects that require less specific elements.

The Struggle with Cloud-Based Models

The dynamic nature of cloud-based models, although beneficial, impacted the reproducibility of my client work. 

This led me to rely on local and open-source software, ensuring a more dependable and consistent workflow.