Baluja, A. (2024). Text is not all you need: Multimodal prompting helps LLMs understand humor. Accepted at Workshop on Computational Humor at COLING 2025.
Including TTS-generated audio in the prompt helps LLMs generate accurate pun explanations.
[arxiv link]
Baluja, S., Marwood, D., & Baluja, A. (2024). Making images from images: Tightly constrained parallel denoising. In Computer Vision – ECCV 2024 Workshops: AI for Visual Arts Workshop and Challenges.
Generate new images out of tiles from existing images by iteratively alternating between diffusion-based denoising and matching tiles between the generated and source images. Works in both latent and pixel space.
[arXiv link] [poster link]
Ebert, T., Baluja, A., Hall, G. N., & Trosseille, C. (2024). Modeling of measurement uncertainties for X-ray radiograph analysis. [Conference presentation]. 25th Topical Conference on High Temperature Plasma Diagnostics.
Recover the range of possible 3D shapes of the capsules used in inertial-confinement-fusion as they implode, via gradient descent.
[abstract link] [poster link]
Baluja, A., Ebert, T., & Hall, G. N. (in preparation). Modeling measurement uncertainties for x-ray radiograph analysis using gradient descent with PEREGRINE. To be submitted to Computer Physics Communications.
By providing a library of simulations of common measurement errors/uncertainties and backpropagating through them, we can explicitly account for them.