My reading list.

. We are hiring Research Interns at @DolbyIn!.

Kapil Dev.

The docking scene.

It’s been a long journey full of challenges, growth, and learning. Thanks to everyone who’s been part of it. ❤️

Turns out years of caffeine, code, and existential dread do lead somewhere. #PhDone

Vibe coding your way into deployment and production seems chill until your app gets cyber attacked -- sometimes a little seriousness in your code is the best vibe!

In a world where the literature is overflowing with diffusion and other large-scale models, it soothes the heart to receive papers for review where efforts are devoted to the theoretical comprehension of VAEs, and the relationship between priors and posteriors!.

Farewell, student life! 🎓 After years of research and ramen, I’m stepping into the real world as a Senior AI Researcher @DolbyIn.

10/ Come see our talk at #BMVC2024 to learn how RISSOLE combines the best of compact model design and state-of-the-art image quality. More details here:

9/ Beyond images, this approach could be extended to video synthesis, multi-modal data, and more. It’s a step toward democratizing access to powerful generative models.

8/ Why it matters:.RISSOLE opens up new possibilities for deploying high-quality generative models on resource-constrained devices—ideal for applications in medical imaging, creative AI, and edge devices.

7/ For example:.•FID on CelebA: RISSOLE = 9.82, RDM = 32.36 (lower is better). •RISSOLE even outperforms variants that use positional encodings or prior blocks for conditioning.

6/ Tested on challenging datasets like CelebA and ImageNet100, RISSOLE consistently outperforms baselines like RDM and patch-based models. We achieve better FID scores while using less memory and computational power!.

5/ Why RISSOLE?.🖼️ Coherent Image Blocks: Blocks blend seamlessly without artifacts. 💡 Compact Model Size: Operates with far fewer parameters. ⚡ Efficient Training: Parallelizes across blocks for speed-ups.

4/ How does RISSOLE work?.•Each block is conditioned on corresponding blocks retrieved from an external database of relevant images. •These retrieved neighbors guide both training and sampling to maintain spatial and semantic consistency.

3/ Key Idea:.Instead of generating the entire image at once, we generate smaller blocks independently. To ensure coherence between blocks, we leverage a novel retrieval-augmented generation (RAG) approach.

2/ Introducing RISSOLE: ParameteR-efficIent DiffuSion Models via Block-wiSe GeneratiOn and RetrievaL-GuidancE! We rethink the design of diffusion models to significantly reduce parameters while retaining exceptional generation quality.

1/ Generative models like diffusion models have incredible power, but their size & computational needs are massive. Training from scratch can take thousands of GPU hours, making them impractical for resource-constrained settings.

🚀 Excited to present our paper RISSOLE at #BMVC2024! Here’s a thread summarizing how we tackle the challenge of making diffusion models both parameter-efficient and high-quality. 🧵