Generative Vision Atlas

Design axis

Representation

What space does the model actually generate in?

Continuous tokens

Per-position continuous vectors generated autoregressively, each sampled by a small diffusion loss instead of a softmax over a discrete codebook.

Deep-compression latent

A latent pushed to much higher spatial compression ratios (up to 128x) than standard VAEs, via residual autoencoding and staged high-resolution adaptation.

Frozen encoder + trained decoder

Keep a pretrained vision encoder entirely frozen as the latent space; train only a lightweight decoder to map that space back to pixels.

Hybrid semantic + detail codebook

Two coupled codebooks/branches: one capturing semantic content, one capturing pixel-level detail, kept aligned so both understanding and generation/reconstruction stay strong.

Multi-scale (next-scale) tokens

A coarse-to-fine pyramid of discrete tokens generated scale by scale, rather than a single flat token grid.

Pixels

Generate directly on raw image pixels; no learned tokenizer or latent space at all.

Semantic / foundation latent

A latent taken from (or aligned to) a pretrained vision foundation encoder's feature space, so it is high-dimensional and semantically structured rather than reconstruction-only.

VAE latent

A compact continuous latent from a KL-regularized autoencoder (e.g. SD-VAE, 4-16 channels), trained purely for pixel reconstruction.