DecQ: Detail-Condensing Queries for Enhanced Reconstruction and Generation in Representation Autoencoders

Tianhang Wang, Yitong Chen, Wei Song, Zuxuan Wu, Min Li, Jiaqi Wang

arXiv:2605.22777 · 2026-05-20 공개 · arXiv · PDF

image-editing latent-diffusion dino-v2 psnr fid condenser-module representation-autoencoder vision-foundation-model

Abstract

Representation Autoencoders (RAEs) leverage frozen vision foundation models (VFMs) as tokenizer encoders, providing robust high-level representations that facilitate fast convergence and high-quality generation in latent diffusion models. However, freezing the VFM inherently constrains its spatial reconstruction capacity, limiting fine-grained generation and image editing; in contrast, incorporating reconstruction-oriented signals via fine-tuning disrupts the pretrained semantic space and degrades generative fidelity. To address this trade-off, we propose DecQ, a simple yet effective framework for RAEs. Specifically, DecQ introduces lightweight detail-condensing queries that extract fine-grained information from intermediate VFM features through condenser modules. These queries are incorporated into the decoder to support reconstruction and are jointly generated with patch tokens during generative modeling. By aggregating information from both shallow and deep layers, DecQ effectively mitigates the reconstruction--generation trade-off, improving both reconstruction quality and generative performance. Our experiments demonstrate that: (1) with only 8 additional queries and 3.9% extra computation, DecQ improves reconstruction over the frozen DINOv2-based RAE, increasing PSNR from 19.13 dB to 22.76 dB; and (2) for generative modeling, DecQ achieves 3.3$\times$ faster convergence than RAE, attaining an FID of 1.41 without guidance and 1.05 with guidance.

한국어 요약

📋 한 줄 요약

**[Representation Autoencoder / Latent Diffusion]** DecQ가 8개 detail-condensing query·3.9% 추가 연산으로 frozen DINOv2 RAE의 reconstruction-generation trade-off 해소 — PSNR 19.13→22.76 dB, 3.3× 빠른 수렴, FID 1.41 (no guidance) / 1.05 (guidance).

🎯 핵심 기여도

💡 핵심 아이디어

RAE의 reconstruction-generation 딜레마는 VFM freeze 자체가 아니라 fine-grained 정보 접근 부재에서 비롯

💡 핵심 아이디어

RAE의 reconstruction-generation 딜레마는 VFM freeze 자체가 아니라 fine-grained 정보 접근 부재에서 비롯되며, condenser로 중간 layer feature를 압축한 query를 decoder에 주입하면서 generative process에 jointly 생성하면 semantic space 보존과 detail 복원을 동시 달성할 수 있다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: VFM 기반 latent generation의 핵심 trade-off를 minimal 비용으로 해소, condenser-query 디자인의 일반 적용 가능 패턴, latent diffusion의 detail 표현력 향상으로 fine-grained 생성·편집 가능성 확장. **한계**: VFM 의존(DINOv2 외 일반화 추

💭 의의 및 한계

**의의**: VFM 기반 latent generation의 핵심 trade-off를 minimal 비용으로 해소, condenser-query 디자인의 일반 적용 가능 패턴, latent diffusion의 detail 표현력 향상으로 fine-grained 생성·편집 가능성 확장. **한계**: VFM 의존(DINOv2 외 일반화 추가 검증), 8 query의 capacity 한계, query 수·condenser 구조의 hyperparameter 의존성.

🚀 실용적 활용