Generative Vision Atlas

strong-followup

Improving Reconstruction of Representation Autoencoder

Siyu Liu, Chujie Qin, Hubery Yin · 2026-02 · arXiv:2602.08620

Why it matters

Augments RAE's semantic features with the low-level colour and texture information a frozen semantic encoder discards, and separately hardens the decoder against latent perturbation through fine-tuning with noise injection — a fix for the artifact sensitivity that high-dimensional semantic latents introduce.

What this paper does

read: full text

Before this

RAE-style tokenizers give diffusion models strong, fast-converging semantic latents from a frozen vision foundation model, but reconstruction fidelity from those latents lags, and the paper frames this reconstruction gap as the primary bottleneck to scaling LDMs further; naively fine-tuning the VFM with combined reconstruction and alignment losses fails because the semantic base manifold keeps shifting during training, leaving low-level detail with no stable reference.

The problem

VFM semantic features lack low-level information such as color and texture, degrading reconstruction fidelity, and separately, the high-dimensional, information-rich latents that result make decoders hypersensitive to small perturbations, producing severe artifacts when decoding imperfect generated latents rather than clean encoded ones.

The idea

Model the data manifold as a smooth semantic base, supplied by the frozen VFM, plus local low-level variation that the VFM ignores; learn only that residual with a small trained encoder and add it to the frozen semantic feature, then separately harden the decoder against latent perturbation through noise-injected fine-tuning.

How it works

A lightweight 6-layer, 768-dimensional trained encoder takes the pixel input, projects it to patch tokens, concatenates them with the frozen VFM's semantic features, and predicts low-level residual tokens r; the final latent is z = LayerNorm(r + u), with u the frozen semantic feature, so the VFM itself never updates. Stage 1 trains a 12-layer decoder with L1 and LPIPS reconstruction loss plus an alignment regularizer (weight eta=5) that keeps the output grounded in the semantic manifold. A toy analysis attributes decoder fragility to excessively large Jacobian magnitudes along directions off the data manifold. Stage 2 fine-tunes the decoder with stochastic noise injected into the latent during training (z-tilde = z + sigma*epsilon, sigma ~ U(0, tau), tau=0.2) plus an adaptive reconstruction-and-GAN loss; at inference, controlled noise is likewise injected into generated latents before decoding to suppress off-manifold artifacts.

Evidence

Reconstruction PSNR reaches about 32.32 on ImageNet, ahead of FLUX-VAE's 31.01 and SVG's 21.87, with SSIM 0.941, LPIPS 0.015, and CKNNA semantic alignment about 0.987 against a fine-tuned-VFM ablation's 0.925 (PSNR 30.88). On class-conditional ImageNet 256x256 generation, gFID reaches 3.77 without guidance (DiT-XL, 400 epochs); with guidance, gFID 2.42 (DiTDH-XL, 800 epochs), with gFDD 58.2 against VA-VAE's 74.3 and SVG's 130.4. On latents corrupted with z + 0.2*epsilon, the original decoder's PSNR collapses to 13.68, while the noise-fine-tuned decoder holds at 21.61.

Limitations

The paper does not quantify the compute overhead of its two-stage training pipeline. The optimal inference noise scale (sigma-bar about 0.08) is manually selected and architecture- dependent. The authors concede that fine-tuning the decoder with a fixed rather than randomly sampled noise scale severely harms reconstruction on clean latents, so the randomized-noise recipe is load-bearing rather than incidental.

Why it matters

It gives an independent confirmation that RAE's reconstruction bottleneck can be closed with a small additive trained residual rather than by unfreezing the encoder, and it isolates a separate failure mode, decoder sensitivity to off-manifold latent perturbation, that traces generation-time artifacts to a cause distinct from reconstruction-time fidelity.

Abstract, in the authors' own words

Recent work leverages Vision Foundation Models as image encoders to boost the generative performance of latent diffusion models (LDMs), as their semantic feature distributions are easy to learn. However, such semantic features often lack low-level information (\eg, color and texture), leading to degraded reconstruction fidelity, which has emerged as a primary bottleneck in further scaling LDMs. To address this limitation, we propose LV-RAE, a representation autoencoder that augments semantic features with missing low-level information, enabling high-fidelity reconstruction while remaining highly aligned with the semantic distribution. We further observe that the resulting high-dimensional, information-rich latent make decoders sensitive to latent perturbations, causing severe artifacts when decoding generated latent and consequently degrading generation quality. Our analysis suggests that this sensitivity primarily stems from excessive decoder responses along directions off the data manifold. Building on these insights, we propose fine-tuning the decoder to increase its robustness and smoothing the generated latent via controlled noise injection, thereby enhancing generation quality. Experiments demonstrate that LV-RAE significantly improves reconstruction fidelity while preserving the semantic abstraction and achieving strong generative quality. Our code is available at https://github.com/modyu-liu/LVRAE.

Research line

Semantic-plus-detail hybridsemerging

Design-axis choices

Problem

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.