AI Paper Atlas

Research map / Image generation / Autoregressive and latent generation

USP: Unified Self-Supervised Pretraining for Image Generation and Understanding

Xiangxiang Chu, Renda Li, Yong Wang

ICCV 2025 · Image generation › Autoregressive and latent generation · approach: Latent diffusion with learned tokenizers · 19 citations (OpenAlex)

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 changedWhereMeasuredEffect
Initialise with USP-pretrained weightsDiT-B/2 400K steps · ImageNet 256 generationFID: 42.62 -> 28.26Improved
Initialise with USP-pretrained weightsDiT-XL/2 400K steps · ImageNet 256 generationFID: 19.94 -> 10.28Improved
Initialise with USP-pretrained weightsSiT-XL/2 400K steps · ImageNet 256 generationFID: 16.97 -> 7.38Improved
Pretrain 1600 instead of 800 epochs (ablation)USP DiT-XL/2 · ImageNet 256 generationFID: 10.28 -> 9.73Improved
Combine USP with REPASiT 400K steps · ImageNet 256 generationFID: 7.9 -> 6.26Improved
Initialise with MAE weights instead of USP (ablation)DiT-B/2 · ImageNet 256 generationFID: 42.62 -> 43.55Hurt
Add pretrained LayerNorm parameters (ablation)USP DiT-B/2 · ImageNet 256 generationFID: 28.76 -> 28.26Improved
Add per-patch pixel normalization target (ablation)USP · ImageNet 256 generationFID: 30.65 -> 28.26Improved
Use VAE normalization instead of ImageNet normalization (ablation)USP · ImageNet 256 generationFID: 78.23 -> 28.26Improved
Keep pretrained weights in last layers instead of reinitialising 6 (ablation)USP DiT-B/2 · ImageNet 256 generationFID: 37.58 -> 28.26Improved
Pretrain with MAE-style masking, not noisy latents (ablation)USP · ImageNet 256 generationFID: 32.20 -> 28.26Improved
Use SFT classifier init instead of USP pretraining (ablation)DiT-B/2 · ImageNet 256 generationFID: 28.26 -> 39.35Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy