Help me find optimal hyper-parameters for Ultimate Stable Diffusion Upscale and complete my masters degree!
A master's student is crowdsourcing data to find the best settings for high-quality AI image upscaling.
A master's student is turning to the public to solve a complex AI optimization problem. Jacob Waters, pursuing an MS in Data Science and AI, is conducting a novel research project to identify the ideal hyperparameters for the 'Ultimate Stable Diffusion Upscaler,' a tool used to enhance image resolution with AI. His study focuses on three key settings: denoise strength, ControlNet strength, and step count, specifically for the domain of creating high-quality, printable oil painting reproductions. To gather robust data, Waters has designed a survey that presents participants with pairwise comparisons of images generated using different hyperparameter combinations.
Participants are asked to judge the upscaled images across three critical categories: fidelity to the original, aesthetic 'prettiness,' and the quality of fine details. The research employs an ELO rating system—commonly used in chess—to rank each hyperparameter configuration in real-time. A public dashboard allows anyone to watch the live rankings evolve as more survey data is collected. Waters has committed to publishing the final results as open-source data, which could provide artists and developers with a validated, optimal starting point for upscaling workflows, saving significant time and computational experimentation.
- Research targets three key Stable Diffusion hyperparameters: denoise strength, ControlNet strength, and step count for upscaling.
- Uses a public pairwise comparison survey, judging images on fidelity, prettiness, and detail for print-quality art.
- Features a live, public ELO ranking graph to show real-time results, with all findings to be released open-source.
Why It Matters
Provides empirical, crowd-validated settings that could standardize and improve quality for AI artists and developers working on image upscaling.