Research map / Image generation / Autoregressive and latent generation
ArcDAE: Asymmetric Rectified Contrastive Diffusion Autoencoder for Unified Representation Learning
Ge Gao, Di Xiong, Zeke Xie, Jian Yang, Shuo Chen
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add the EMA teacher anchor to timestep-aware progressive augmentation (ablation) | DBAE + TAPA · FFHQ | linear-probe AP: 0.125 -> 0.679 | Improved |
| replace static weak augmentation with dynamic TAPA in ASR (ablation) | ArcDAE · FFHQ | AP: 0.666 -> 0.679 | Improved |
| add timestep-aware adaptive weighting (TAAW) to TGA (ablation) | ArcDAE · FFHQ | FID: 10.85 -> 10.28 | Improved |
| add TGA (dual granularity alignment) (ablation) | ArcDAE · FFHQ / CelebA | LPIPS: 0.02 lower LPIPS with TGA | Improved |
| add ArcDAE rectification to DBAE | DBAE · FFHQ linear probe | AP: 0.664 -> 0.681 | Improved |
| raise alignment weight lambda_align from 0.1 to 1.0 (ablation) | ArcDAE · FFHQ (128x128) | AP: 0.681 -> 0.620 | Hurt |
| lower lambda_align from 0.1 to 0.01 (ablation) | ArcDAE · CelebA (64x64) | AP: 0.642 -> 0.676 on CelebA; 0.681 -> 0.665 on FFHQ | Mixed |
| use ArcDAE instead of DBAE-d for reconstruction | ArcDAE · CelebA-HQ | MSE: 2.49 × 10−3 -> 2.79 × 10−3 (LPIPS 0.072 -> 0.065) | Hurt |
| add total-correlation regularization to ArcDAE | ArcDAE+TC · CelebA | TAD: 0.210 -> 0.385 | Improved |
| replace DBAE with ArcDAE for unconditional generation | ArcDAE · FFHQ | FID: 11.25 -> 10.28 | Improved |
| compare ArcDAE FID with DDPM | ArcDAE · FFHQ unconditional | FID: 9.14 (DDPM) vs 10.28 | Hurt |
Closest papers by meaning
- DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence (NeurIPS 2026 · similarity 0.88)
- Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think (NeurIPS 2025 · similarity 0.87)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.87)
- Diffusion Bridge AutoEncoders for Unsupervised Representation Learning (ICLR 2025 · similarity 0.87)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.87)
- Steering Optimisation Trajectories in Diffusion Representation Learning (NeurIPS 2026 · similarity 0.86)
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.
