Avataar AI
3D reconstruction & novel-view synthesis
Sparse-image-conditioned novel-view synthesis for 3D reconstruction at Avataar AI.
Fine-tuned sparse-image-conditioned diffusion models and worked on conditioning, distributed training and SLURM, building on a senior engineer’s initial setup.
Scope
Worked on sparse-view-conditioned novel-view synthesis during roughly the first two months at Avataar AI, continuing an existing project whose initial setup had been established by a senior engineer.
Model work
Fine-tuned a Stable Diffusion model in a Zero123-like setting to generate novel views conditioned on sparse input images. Worked on changes to how images condition the diffusion model.
Training infrastructure
Handled distributed training and SLURM job submission on the HPC cluster. Used its full 4 × 8 H100 capacity (32 GPUs) for a few days.
Tools
Stable Diffusion · Image conditioning · Distributed training · SLURM · HPC