Research map / Image generation / Autoregressive and latent generation
Diffusion Autoencoders: Toward a Meaningful and Decodable Representation
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Condition the DDIM decoder on a learned semantic code and retain an encoded stochastic code | DDIM autoencoding · CelebA-HQ reconstruction | SSIM: 0.917 -> 0.991 | Helped |
| Condition DDIM on the learned semantic code | DDIM unconditional generation · FFHQ 128, T=10 | FID: 29.56 -> 20.80 | Helped |
| Condition DDIM on the learned semantic code | DDIM unconditional generation · Horse 128, T=100 | FID: 5.97 -> 6.71 | Hurt |
| Use a learned semantic code for few-shot rejection sampling | D2C-style conditional generation · CelebA Male, binary | FID: 13.44 -> 11.52 ± 1.19 | Helped |
| Use a learned semantic code for few-shot rejection sampling | D2C-style conditional generation · CelebA Male, positive-unlabeled | FID: 16.39 -> 9.54 ± 0.54 | Helped |
| Encode x_T from the input rather than sample it during decoding (ablation) | Diffusion Autoencoder · CelebA-HQ reconstruction | SSIM: 0.677 -> 0.991 | Helped |
| Increase semantic code size from 64D to 128D (ablation) | Diffusion Autoencoder without encoded x_T · CelebA-HQ reconstruction, 48M training images | SSIM: 0.521 -> 0.582 | Helped |
| Increase semantic code size from 128D to 256D (ablation) | Diffusion Autoencoder without encoded x_T · CelebA-HQ reconstruction, 48M training images | SSIM: 0.582 -> 0.606 | Helped |
| Increase semantic code size from 256D to 512D (ablation) | Diffusion Autoencoder without encoded x_T · CelebA-HQ reconstruction, 48M training images | SSIM: 0.606 -> 0.631 | Helped |
| Replace linear beta schedule with constant 0.008 (ablation) | Latent DDIM · LSUN Horse 128 validation | FID: 13.36 -> 10.50 | Helped |
| Replace MSE with L1 noise-prediction loss (ablation) | Latent DDIM · LSUN Horse 128 validation | FID: 13.36 -> 11.65 | Helped |
| Use a 20-layer latent MLP rather than the 10-layer, size-2048 reference configuration (ablation) | Latent DDIM · LSUN Horse 128 validation | FID: 13.36 -> 9.30 | Helped |
Closest papers by meaning
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.90)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.90)
- Deconstructing Denoising Diffusion Models for Self-Supervised Learning (ICLR 2025 · similarity 0.90)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.90)
- DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence (NeurIPS 2026 · similarity 0.89)
- Bridging Data Gaps in Diffusion Models with Adversarial Noise-Based Transfer Learning (ICML 2024 · similarity 0.89)
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.