AI Paper Atlas

Research map / Image generation / Autoregressive and latent generation

A Flat Vocabulary or a Rich Hierarchy? Re-introducing Intrinsic Structure Transforms the Autoregressive Image Generation

Landis He, Shikang Zheng

ICML 2026 · Image generation › Autoregressive and latent generation · approach: Next-scale and token autoregressive images

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add MASC hierarchical vocabulary reduction (k=8,192)LlamaGen-XL · ImageNet 256x256 class-conditionalFID: 2.87 -> 2.58Improved
add k-means clustering prior instead of MASC (ablation)LlamaGen-XL · ImageNet 256x256 class-conditionalFID: 2.87 -> 3.09Hurt
add k-means prior to LlamaGen-L (ablation)LlamaGen-L · ImageNet 256x256 class-conditionalFID: 3.38 -> 3.81Hurt
add MASC to LlamaGen-LLlamaGen-L · ImageNet 256x256 class-conditionalFID: 3.38 -> 2.92Improved
reduce MASC vocabulary from 8,192 to 4,096 (ablation)LlamaGen-XL · ImageNet 256x256 class-conditionalFID: 2.58 -> 2.49Improved
reduce MASC vocabulary from 4,096 to 2,048 (ablation)LlamaGen-XL · ImageNet 256x256 class-conditionalFID: 2.49 -> 2.65Hurt
use random clustering of the vocabulary (ablation)VAR-d24 · ImageNet 256x256 class-conditionalFID: 2.09 -> 3.84Hurt
add MASC (2k) to VAR-d24VAR-d24 · ImageNet 256x256 class-conditionalFID: 2.09 -> 1.98Improved
add MASC (4k) to RAR-L with a large-vocabulary tokenizerRAR-L + LlamaGen · ImageNet 256x256 class-conditionalconvergence and FID: diverged -> 1.61Improved
use MASC target space instead of random or k-means at 4k (ablation)RAR-L · ImageNet 256x256 class-conditionalFID: 4.24 (k-means) -> 1.61Improved
use manifold-aligned distance on top of density-driven construction (full MASC vs centroid) (ablation)LlamaGen-XL · ImageNet 256x256 class-conditionalFID: 3.02 -> 2.58Improved
use density-driven bottom-up construction vs k-means++ (ablation)LlamaGen-XL · ImageNet 256x256 class-conditionalFID: 3.17 -> 3.02Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy