Generative Vision Atlas

strong-followup

dRAE: Representation Autoencoder with Hyper-Spherical Codes

Tianren Ma · 2026-07 · arXiv:2607.22148

Why it matters

Discretizes the continuous semantic latent that representation autoencoders produce, using hyper-spherical quantization to avoid the codebook collapse that Euclidean quantization suffers on anisotropic features, and scales to a 131,072-entry codebook. The bridge between the representation-latent line and discrete-token generation.

What this paper does

read: full text

Before this

RAE established that a frozen vision-foundation-model encoder paired with a trained decoder yields a continuous semantic latent that is strong for both reconstruction and flow-matching generation. Discretizing that continuous space with ordinary vector quantization, as a prior VQRAE attempt did, hit codebook collapse and stopped improving once the vocabulary grew past roughly 16,384 codes.

The problem

Standard Euclidean-distance codebook assignment is mismatched with the geometry of pretrained representation features, where semantic content lives mainly in feature orientation rather than magnitude, so quantization saturates and cannot scale to large vocabularies while staying semantically coherent.

The idea

Discretize RAE latents with Hyper-Spherical Quantization, which assigns each token to the codebook entry with the highest cosine similarity to it rather than the smallest Euclidean distance, decoupling semantic routing from feature magnitude so the codebook can grow without collapsing.

How it works

A frozen SigLIP2 ViT-So400M encoder supplies the continuous representation, exactly as in plain RAE. Hyper-Spherical Quantization then routes each token to the codebook entry maximizing angular similarity between token and code directions, while a separate Euclidean commitment loss preserves magnitude information needed for reconstruction; the nearest-neighbor search is spherical, but the retrieved code vector itself is unnormalized. A trained decoder, guided in part by a self-distillation teacher, reconstructs images from the resulting discrete codes. Downstream generation is a discrete diffusion model trained with a time-weighted masked cross-entropy loss over the token sequence.

Evidence

On ImageNet-1K reconstruction, dRAE reaches rFID 0.69 with a 16,384-entry codebook and rFID 0.42 with a 131,072-entry codebook, improving as vocabulary grows, with over 90 percent codebook utilization and roughly 12,000 codes active per forward pass against roughly 6,000 for standard VQ at comparable size. Class-conditional generation at a 65,536-entry codebook reaches gFID 4.45 with Inception Score 287.3. Paired with a 7B multimodal LLM on the same SigLIP2 encoder at 512 resolution, dRAE reaches MMBench 81.5, TextVQA 67.7, GQA 62.4, and SEED 72.7, and text-to-image generation trained on 12M pairs reaches GenEval 0.63 and DPG-Bench 80.58.

Limitations

The paper concedes reconstructing high-dimensional patch embeddings is harder than reconstructing raw pixel values, attributing this to intrinsic noise in the encoder's own representations and the absence of perceptual supervision at that stage. It also reports that imposing a strict unit-sphere prior on the inputs themselves, rather than restricting spherical geometry to code assignment alone, degrades reconstruction, dropping PSNR from 22.5 to 20.6.

Why it matters

The d in dRAE stands for discrete. What it changes relative to plain RAE is specifically a discretization stage built around a hyper-spherical codebook designed for anisotropic feature geometry, not a new decoder architecture or training curriculum in the sense of reordered training stages, though the decoder is retrained around the new discrete codes. This gives the representation-latent line a genuine bridge to language-model-style next-token generation and multimodal LLM training, which continuous RAE latents cannot directly support.

Abstract, in the authors' own words

In this work, we aim to discretize the high-dimensional visual representations to bridge the gap with language models - a non-trivial challenge, as existing quantization methods suffer from codebook collapse, failing to scale while preserving semantic coherence. We identify the root cause as metric mismatch: standard Euclidean codebook objectives are fundamentally misaligned with the anisotropic geometry of representation space, leading to codebook embeddings with high-variance magnitude scales and uneven angular distributions that hinder scalability. To address this, we propose Hyper-Spherical Quantization (HSQ), which decouples semantic content from feature magnitude via angular routing, preventing code assignment from being dominated by scale rather than meaning. The resulting discrete Representation Autoencoder (dRAE) achieves high-fidelity reconstruction while preserving semantic integrity and supporting scalable codebook budget. Extensive experiments demonstrate consistent performance gains as the vocabulary size scales to 131{,}072, along with 100\% codebook utilization, simplified training pipeline, and strong performance across understanding and generation tasks.

Research lines

Representation-space generationcontestedDiscrete-token autoregressioncontested

Builds on

Nothing recorded yet.

Built on by

Nothing recorded yet.

This is a strong-followup entry — a short-form summary. Full paper-page explanations (before/problem/ core idea/representation/architecture/objective/conditioning/training/inference/results/ ablations/limitations) are written for landmark and core papers first; see PROJECT_STATE.md for the schedule.