DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence
DC-SAE compresses images 32x and speeds diffusion training while beating DC-AE on ImageNet.
DC-SAE is a decoupled semantic autoencoder intended to combine high image compression with faster diffusion convergence. It uses semantic encoders for compact latents and a pixel-level encoder to retain reconstruction detail. On ImageNet at 512 by 512, it achieves 32x spatial compression, 29.79 PSNR, and 3.37 gFID, improving on DC-AE by 13.5% and 54.9% respectively. A 1.6B-parameter DiT using DC-SAE scores 0.84 on GenEval and 86.007 on DPG-Bench for 1024 text-to-image generation.
- DC-SAE combines a semantic encoder with a pixel-detail encoder.
- ImageNet 512 results are 32x compression, 29.79 PSNR, and 3.37 gFID.
- PSNR and gFID improve 13.5% and 54.9% over DC-AE.
- A 1.6B DiT scores 0.84 GenEval and 86.007 DPG-Bench at 1024.
- AI models
- DiT
From the original article
1 min read at huggingface.co
High-compression tokenizers are essential for scaling latent image generative models. However, aggressive compression creates a fundamental tradeoff between reconstruction fidelity and generation efficiency: high compression image encoder always increases the learning difficulty of diffusion training, resulting in slow model convergence.
By Xu Huang, Ye Huang, Zijun Liao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.