ControlNet Scaling Laws

Kapatsyn, Diana

Домівка
→
Students Research & Project Works | Роботи студентів
→
Факультет прикладних наук
→
Освітня програма наук про дані
→
2024
→
Перегляд матеріалів

dc.contributor.author	Kapatsyn, Diana
dc.date.accessioned	2024-08-26T12:14:23Z
dc.date.available	2024-08-26T12:14:23Z
dc.date.issued	2024
dc.identifier.citation	Kapatsyn Diana. ControlNet Scaling Laws. Ukrainian Catholic University, Faculty of Applied Sciences, Department of Computer Sciences. Lviv 2024, 51 p.	uk
dc.identifier.uri	https://er.ucu.edu.ua/handle/1/4684
dc.language.iso	en	uk
dc.subject	Scaling Laws	uk
dc.subject	ControlNet Scaling Laws	uk
dc.subject	resource allocation	uk
dc.title	ControlNet Scaling Laws	uk
dc.type	Preprint	uk
dc.status	Публікується вперше	uk
dc.description.abstracten	Effectively training large-scale deep learning models is costly, and requires careful planning and resource allocation. One strategy involves fitting simple parametric functions, such as logarithmic functions, on smaller-scale experiments and extrap- olating them to predict model performance and associated costs. These empirical "scaling laws" are then used to predict the required resources for achieving a given level of performance. This approach is widely used for Large Language Models but is insufficiently investigated for computer vision generative models. Today, diffu- sion models dominate image generation and ControlNet is one of the most popular ways to customize and control them. This work makes three contributions. First, we have estimated the scaling laws for ControlNet quality depending on the dataset size. Second, we have shown that task-specific metrics, such as edge detection metrics for Canny edges are more suit- able for predicting image quality compared to the ControlNet training and valida- tion loss itself. Finally, we present a practical recommendations for dataset size for ControlNet training. The code an and data are available on GitHub1 and HuggingFace respectively.	uk