Research map / Image generation / Autoregressive and latent generation
USP: Unified Self-Supervised Pretraining for Image Generation and Understanding
Xiangxiang Chu, Renda Li, Yong Wang
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 |
|---|---|---|---|
| Initialise with USP-pretrained weights | DiT-B/2 400K steps · ImageNet 256 generation | FID: 42.62 -> 28.26 | Improved |
| Initialise with USP-pretrained weights | DiT-XL/2 400K steps · ImageNet 256 generation | FID: 19.94 -> 10.28 | Improved |
| Initialise with USP-pretrained weights | SiT-XL/2 400K steps · ImageNet 256 generation | FID: 16.97 -> 7.38 | Improved |
| Pretrain 1600 instead of 800 epochs (ablation) | USP DiT-XL/2 · ImageNet 256 generation | FID: 10.28 -> 9.73 | Improved |
| Combine USP with REPA | SiT 400K steps · ImageNet 256 generation | FID: 7.9 -> 6.26 | Improved |
| Initialise with MAE weights instead of USP (ablation) | DiT-B/2 · ImageNet 256 generation | FID: 42.62 -> 43.55 | Hurt |
| Add pretrained LayerNorm parameters (ablation) | USP DiT-B/2 · ImageNet 256 generation | FID: 28.76 -> 28.26 | Improved |
| Add per-patch pixel normalization target (ablation) | USP · ImageNet 256 generation | FID: 30.65 -> 28.26 | Improved |
| Use VAE normalization instead of ImageNet normalization (ablation) | USP · ImageNet 256 generation | FID: 78.23 -> 28.26 | Improved |
| Keep pretrained weights in last layers instead of reinitialising 6 (ablation) | USP DiT-B/2 · ImageNet 256 generation | FID: 37.58 -> 28.26 | Improved |
| Pretrain with MAE-style masking, not noisy latents (ablation) | USP · ImageNet 256 generation | FID: 32.20 -> 28.26 | Improved |
| Use SFT classifier init instead of USP pretraining (ablation) | DiT-B/2 · ImageNet 256 generation | FID: 28.26 -> 39.35 | Hurt |
Closest papers by meaning
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.89)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.88)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.88)
- Nested Diffusion Models Using Hierarchical Latent Priors (CVPR 2025 · similarity 0.88)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.88)
- On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · 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.
