Project Info
Inspiration
We saw AppLovin’s challenge as an opportunity to help Axon understand what makes a creative work — not just who to show it to. Our team loves multimodal modeling and wanted to bring more interpretable intelligence into real-time ad recommendations.
What it does
Our system converts each ad (image or video) into a latent embedding and projects it onto five human-interpretable creative style axes: wealthy, limited-offer, calm, honest, certified Each ad gets a 5-dimensional vector: \( \mathbf{f}(x) = \left( \langle \hat{z}x, \hat{w}{\text{wealthy}} \rangle,, \langle \hat{z}x, \hat{w}{\text{limited}} \rangle,, \ldots \right) \) capturing how strongly that creative expresses each concept.
How we built it
Use ImageBind to embed each image/video: (\mathbf{z} \in \mathbb{R}^{1024}) Perform PCA on all creatives to find high-variance semantic directions. Select interpretable axes aligned with principal components using text embedding similarity. Normalize embeddings and extract cosine-based activations: \( \text{feature}_i(x) = \frac{\mathbf{z}_x}{|\mathbf{z}_x|} \cdot \frac{\mathbf{w}_i}{|\mathbf{w}_i|} \quad \in [-1, 1] \) This transforms raw creatives into features that are: distinctive, predictive, and scalable to millions.
Challenges we ran into
Few labeled ads → needed unsupervised structure discovery Avoiding meaningless “low-correlation” features that barely activate Ensuring each axis corresponded to recognizable strategy rather than noise Multimodal video + image handling without quality loss
Accomplishments we're proud of
Unified video + image features in one embedding space Found orthogonal and explainable ad attributes Reduced complexity: 1024 → 5 meaningful dimensions Feature activations matched real creative classes we observed qualitatively
What we learned
The best features are not just decorrelated — they must: show up often, align with marketing intuition, and differentiate user-facing persuasion strategies.
What's next
Automatically tag incoming creatives for Axon Extend dimensions using weak supervision + OCR cues Our technical paper which writes up everything in detail is available here: https://drive.google.com/file/d/1hu5jRldgn0yXB-3oFjAGVWi_pctjlQ9t/view?usp=sharing
This project did not link a GitHub repository.
Analysis
No indexed repository for this project, so there are no commit stats to show.
Technology
- PyTorchUnchecked
No repository was indexed for this project, so these Devpost claims have not been checked against code.
AI coding agents
No repository was indexed, so agent usage could not be checked.
Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.
Codebase size
No repository was indexed, so there is no codebase to measure.
This project did not link a GitHub repository, so there is nothing to diagram.
This project did not link a GitHub repository, so its feature claims have not been checked against code.
Export this project's context (description, README, evidence, key source files) to chat with an AI agent elsewhere.