Finetuning Diffusion Model on Bitcoin Dataset
Date Issued
May 2025
Author(s)
Advisor
Abstract
Diffusion models like Stable Diffusion are currently State of the Art for generating images from text
prompts. One of its most impressive capabilities is to synthesize highly realistic, high-quality, and detailed
images that closely align with the given prompt. On a different front, there exist chart-based images that
visually represent the price flow of financial assets over specific time frames. In this work, we focus on
fine-tuning the Stable Diffusion XL (SDXL) model, which is the largest variant in the Stable Diffusion
family that uses the diffuser framework on a custom dataset consisting of Bitcoin price charts and a
description of the price flow. These charts represent various time frames and are labeled with descriptive
prompts. Our goal is to evaluate how effectively SDXL can be adapted to this specialized task, which
differs from the original training objective of the model. Specifically, synthesize a forecast chart based
on the user’s prompt that represents the future trajectory of the asset.
prompts. One of its most impressive capabilities is to synthesize highly realistic, high-quality, and detailed
images that closely align with the given prompt. On a different front, there exist chart-based images that
visually represent the price flow of financial assets over specific time frames. In this work, we focus on
fine-tuning the Stable Diffusion XL (SDXL) model, which is the largest variant in the Stable Diffusion
family that uses the diffuser framework on a custom dataset consisting of Bitcoin price charts and a
description of the price flow. These charts represent various time frames and are labeled with descriptive
prompts. Our goal is to evaluate how effectively SDXL can be adapted to this specialized task, which
differs from the original training objective of the model. Specifically, synthesize a forecast chart based
on the user’s prompt that represents the future trajectory of the asset.
File(s)![Thumbnail Image]()
Name
AndreyZharkovFinalMSc-2025-abstract.pdf
Size
141.57 KB
Format
Adobe PDF
Checksum (MD5)
87acda2275aefa364b0a9cd1d03332b3