Research · AI · vision
HELIX-AI
HELIX-IM — Hyperspherical Entropy-Lax Image Modeling
A ViT tokenizer with residual spherical quantization (R-BSQ) and a causal Transformer prior, implemented in Rust with Burn.
Problem
Learned codebooks (VQ-GAN style) can collapse: a few codes hog usage and the representation thins out. Image tokens need a more stable geometry.
Approach
HELIX-IM skips the learned codebook. It works on the unit sphere with multi-level residual spherical quantization (R-BSQ) and a causal Transformer prior. Inspired by ViT-VQGAN / VIM, with crates helix-math, helix-model, helix-data, helix-train, and helix-cli.
Outcome
An open-source research implementation. We do not report FID or throughput numbers that are not in the repo; the deliverable is the architecture and the code.
Stack
Rust · Burn · ViT · R-BSQ · Transformer
Services