Skip to content

DC-AE - Deep Compression Autoencoder

32x spatial compression autoencoder from MIT Han Lab, core component of SANA. Replaces standard 8x VAE (SD/FLUX) with 4x fewer tokens at any resolution.

Configuration: AE-F32C32P1

  • F32 - 32x spatial downsampling (vs 8x in SD/FLUX)
  • C32 - 32 latent channels (vs 4 in SD)
  • P1 - patch size 1 (no patchification in DiT)

Token Count Comparison

Resolution SD/FLUX (F8, P2) SANA (F32, P1) Reduction
512x512 1024 256 (16x16) 4x
1024x1024 4096 1024 (32x32) 4x
2048x2048 16384 4096 (64x64) 4x
4096x4096 65536 16384 (128x128) 4x

Reconstruction Quality (ImageNet)

Metric DC-AE F32C32 SD VAE F8C4
rFID 0.34 0.31
PSNR 29.29 —
SSIM 0.84 —
LPIPS 0.05 —

Near-identical to 8x VAE quality despite 4x more compression.

Key Techniques

  1. Residual Autoencoding - learns residuals on space-to-channel features, stabilizes training at high compression
  2. Decoupled High-Resolution Adaptation - 3-phase training to avoid generalization penalty on high-res
  3. Tiling support - pipe.vae.enable_tiling(tile_sample_min_height=1024, tile_sample_min_width=1024) enables 4K decode within 22GB VRAM

Variants

  • dc-ae-f32c32-sana-1.0 - original
  • dc-ae-f32c32-sana-1.1 - improved reconstruction
  • dc-ae-lite-f32c32 - lighter, faster inference, smaller memory

Impact on Diffusion

The 4x token reduction compounds with SANA's linear attention O(N): fewer tokens AND linear complexity = orders of magnitude faster at high resolution. At 4K the combined effect makes generation feasible where quadratic attention models cannot run.

Latent Space Properties

  • Scaling factor and shift factor applied during encode/decode (see VAE config)
  • latents_mean and latents_std normalization for stable training
  • BF16 precision recommended for encode/decode

Code

  • GitHub: mit-han-lab/efficientvit (applications/dc_ae/)
  • Diffusers PR: #10510, #10583 (tiling support)
  • HuggingFace: Efficient-Large-Model/dc-ae-f32c32-sana-1.1