cs.CVOct 9, 2026
SaveAPEX: Asynchronous Prefix Denoising for Extreme Compression
Organizations: Lightricks · Hebrew University of Jerusalem
Abstract
Fewer latent tokens promise cheaper diffusion. Yet aggressive latent compression creates a reconstruction-generation dilemma: adding channels improves reconstruction but makes generation harder. We introduce APEX, short for Asynchronous Prefix denoising for EXtreme compression, which pairs a hierarchical latent space with asynchronous denoising over channel groups. Our autoencoder orders channel prefixes from structure to detail, while one shared denoiser generates them at staggered local times without adding tokens. On ImageNet-512, APEX outperforms existing methods at every evaluated channel count under 64x and 128x compression. At matched training and guided sampling compute, it also outperforms a 256-token model at every evaluated budget, with larger gains at lower compute. Code is available at https://daviddinkevich.github.io/APEX.
Figures & tables
Figure 1: APEX enables extreme-compression latent diffusion. (a) An f128c512 latent represents a image with 16 spatial tokens and orders channels from structure to detail. (b) One global clock denoises those levels at staggered local times; is the delay before each later level starts. (c) Ordering generation this way improves the reconstruction–generation trade-off relative to a flat latent.
Figure 2: Channel hierarchy. We divide the channels of each latent token into ordered groups. The cumulative prefix through group defines level . Resolution-supervised autoencoder. We sample a level, zero all later groups, and decode the retained prefix at full resolution. Reconstruction and target are then resized to that level’s assigned resolution before computing the autoencoder loss. Diffusion training. A global clock is sampled from and mapped to one local timestep per group. Unstarted groups are zeroed, and the shared DiT predicts velocities for the denoising groups. Inference. Groups start in sequence along the same clock. Each newly active group is initialized from Gaussian noise, and generation continues until the full latent is clean.
Figure 3: Staged generation. Decodes after each level reaches . Additional samples in Appendix A.1 .
Figure 4: Global clock density.
Figure 5: APEX degrades far more slowly as latent width increases. Left: 64-token f64 latents; right: 16-token f128 latents. At f64, APEX changes from gFID as channels increase , while DC-AE degrades from . At f128, increasing channels from changes APEX from , compared with DC-AE and DC-AE 1.5 . The boxed values summarize performance at f64c128 and f128c512. All runs use matched denoiser-training FLOPs and total NFE.
| Latent | Matched rFID | PSNR | Reference |
|---|---|---|---|
| f64c128 | 0.22 | 26.15 | Official DC-AE |
| f64c256 | 0.18 | 26.8 | Trained by us |
| f64c512 | 0.13 | 27.30 | Trained by us |
| f128c512 | 0.23 | 25.73 | Official DC-AE |
| f128c1024 | 0.20 | 26.1 | Trained by us |
Table 1: Matched full-latent reconstruction quality. Every method at a given latent geometry is trained to the same rFID shown below.
| Configuration | gFID | gFID |
| Build APEX from DC-AE | ||
| DC-AE | 13.27 | |
| APEX levels | 20.37 | |
| cumulative-prefix diffusion training | 17.83 | |
| APEX diffusion | 8.12 | 0.00 |
| Remove components from APEX | ||
Table 2: Building APEX from DC-AE. The first block successively introduces APEX components; the second removes components from the complete method. All runs use f64c128 and SiT-B/1 at matched compute, and gFID is measured relative to APEX.
Figure 6: APEX makes high compression advantageous under constrained compute. (a) Guided gFID versus relative denoiser inference FLOPs, where matches f64 at 250 NFE with f32 at 64 NFE. (b) CFG sweep at the budget. Stars mark the selected guidance scale for each model.
Figure 7: Prefix reconstructions. DC-AE 1.5 (top) vs APEX (bottom) prefixes. Insets show the same crop at 16 channels. High-frequency detail leaks into DC-AE 1.5’s early levels.
Figure 8: Gap, allocation, and level contribution. (a) We vary the stagger gap while using . At , the tail interval is absent and the process reduces to synchronous diffusion. (b) Holding , we vary , the training probability assigned to the interval where is denoising. Uniform clock sampling corresponds to . (c) For the checkpoint, we stop inference when , , or completes and redistribute the same 250 NFEs over the retained clock interval. Generation improves at every endpoint, showing that all three levels contribute.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Latent | Tokens | Batch size | Training steps | Training FLOPs |
|---|---|---|---|---|
| f32c32 | 256 | 1,024 | 250,000 | |
| f64, all widths | 64 | 2,048 | 500,000 | |
| f128, all widths | 16 | 4,096 | 1,000,000 |
Table 3: Main diffusion-training schedules for ImageNet-512. Batch sizes are global.
| Latent | Tokens | Steps | Batch | Training FLOPs | gFID |
|---|---|---|---|---|---|
| SD-VAE f8c4 | 1,024 | 3,000,000 | 256 | 2.62 | |
| APEX f64c | 64 | 1,000,000 | 2,048 | 2.62 |
Table 4: Training compute to match the published f8c4 SiT-XL gFID on ImageNet-512. Batch sizes are global. Training FLOPs use Equation 10 .
Figure 11: Three-level staggered clock with . Before a level starts, its input is zero. During the descending segment it is denoised. Once its local time reaches zero, it stays fixed as context for the remaining levels.