Research map / Image generation / Fast diffusion sampling
SSDD: Single-Step Diffusion Decoder for Efficient Image Tokenization
Théophane Vallaeys, Jakob Verbeek, MATTHIEU CORD
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 SSDD hybrid U-Net/transformer decoder instead of DiTo U-Net (ablation) | DiTo-S-LPIPS · ImageNet 128x128 f8c4 | rFID: 3.17 -> 2.01 (-1.16) | Improved |
| add REPA loss (ablation) | SSDD decoder · ImageNet 128x128 f8c4 | rFID: 2.01 -> 1.58 (-0.44) | Improved |
| add REPA loss (ablation) | SSDD decoder · ImageNet 128x128 f8c4 | PSNR: 23.38 -> 22.82 | Hurt |
| use KL encoder instead of z-norm (ablation) | SSDD decoder · ImageNet 128x128 f8c4 | rFID: 1.58 -> 1.32 | Improved |
| add t-spacing sampler (8 steps) (ablation) | SSDD decoder · ImageNet 128x128 f8c4 | rFID: 1.30 -> 1.25 | Improved |
| add EMA (ablation) | SSDD decoder · ImageNet 128x128 f8c4 | rFID: 1.25 -> 1.17 | Improved |
| add shared encoder (ablation) | SSDD decoder · ImageNet 128x128 f8c4 | rFID: 1.17 -> 1.07 | Improved |
| add shared pre-training (ablation) | SSDD decoder · ImageNet 128x128 f8c4 | rFID: 1.07 -> 0.97 | Improved |
| add single-step distillation (ablation) | SSDD(8)-M · ImageNet 128x128 f8c4 | rFID: 0.97 -> 1.04 | Hurt |
| add GAN loss to GAN-free model (ablation) | SSDD decoder with shared encoder · ImageNet 128x128 f8c4 | rFID: 1.07 -> 1.07 | No clear effect |
| add GAN loss to distilled single-step model (ablation) | SSDD-M · ImageNet 128x128 f8c4 | rFID: 1.04 -> 1.03 | No clear effect |
| add loss-preserving distillation instead of L2-only distillation (ablation) | SSDD-M · ImageNet 128x128 f8c4 | rFID: 1.20 -> 1.04 | Improved |
Closest papers by meaning
- Epsilon-VAE: Denoising as Visual Decoding (ICML 2025 · similarity 0.89)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.88)
- End-to-End Training for Unified Tokenization and Latent Denoising (NeurIPS 2026 · similarity 0.88)
- DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence (NeurIPS 2026 · similarity 0.88)
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.88)
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.87)
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.
