Storytuning

Demo

About

The problem it solves Generative AI models like Midjourney and ChatGPT have revolutionized content creation, but they were trained on massive datasets scraped from the internet—often including artwork, photographs, and writings from creators who never gave their consent. This raises a critical ethical and legal question: who owns a style, and who deserves to benefit when AI learns from it? Today, creators are excluded from this equation. Their work powers the AI industry, yet they receive no recognition, control, or compensation. Storytuning addresses this imbalance by giving creators ownership over both the data they contribute and the models trained on it. Through the platform, creators can upload their artwork, store it on IPFS, and register it on-chain as verifiable intellectual property using Story Protocol. They can then use this registered IP to fine-tune open-source AI models into personalized generators that reflect their unique artistic style. Importantly, the resulting fine-tuned models are not owned by a platform or corporation—they are owned by the creators themselves as on-chain assets. These models can be licensed to other users, and every time a model is used to generate content, the creator automatically receives royalties via programmable revenue sharing. This system enables creators to monetize their data, own the AI trained on their work, and track its usage in a transparent and trustless way. In an age where AI risks replacing human creativity, Storytuning empowers creators to lead the future of generative technology—not be left behind by it.

How it's made

Challenges we ran into One of the main challenges we faced during the hackathon was fine-tuning model performance. Since we're not AI scientists, it was difficult to determine how much data is "enough" or which exact model architecture would yield meaningful results when fine-tuned. There was a steep learning curve in understanding how dataset size, style consistency, and training parameters impact the final output quality of the personalized AI models. Additionally, because our system relies on fine-tuning existing open-source AI models (e.g. Stable Diffusion), the performance of the final output is still heavily dependent on the base model’s pretraining data. We don’t have full visibility into what datasets were used to train these foundational models, which introduces a layer of uncertainty in both performance and the fairness of revenue attribution. For instance, if a model already exhibits strong bias toward a certain aesthetic, it may require far less data to achieve strong fine-tuning results but it’s hard to know who deserves how much credit. We also faced infrastructure limitations. Running live fine-tuning servers with dedicated GPU capacity was not feasible during the hackathon. As a result, we couldn't deliver true real-time fine-tuning as part of our MVP. Instead, we designed a workaround: fine-tuning requests are stored in Firebase, and a Google Colab notebook actively polls for new requests and processes them manually. It’s not ideal for scale, but it allowed us to demonstrate the core functionality of creator-driven model training within the hackathon constraints. Despite these challenges, we built a working system that demonstrates the potential of on-chain IP, user-owned models, and fair compensation for data contribution in the generative AI era.