APEX Compresses Diffusion Latents 128x and Beats a 256-Token Model on ImageNet-512

Researchers have presented APEX on arXiv, a method that targets the main problem of cheap diffusion: the more tightly a latent is compressed, the harder it is to generate. On ImageNet-512, APEX outperforms existing methods at every tested channel count in the 64x and 128x compression regimes. With equal training cost and guided sampling, it also beats a 256-token model at every tested budget, and the lower the compute, the wider the gap. The code is published on the project page.

The dilemma behind the work

The cost of latent diffusion depends directly on the number of tokens the denoiser processes. Fewer tokens mean faster, cheaper training and inference. But the information in an image has to go somewhere. With few tokens, it must be packed into the channels of each token.

This creates a tension the authors call the reconstruction-generation dilemma. More channels let the autoencoder reconstruct the image more accurately. At the same time, such a dense latent is harder for the diffusion model to learn. The gain in reconstruction is eaten up by a drop in generation quality.

What APEX proposes

The name stands for Asynchronous Prefix denoising for EXtreme compression. The solution has two parts that work as a pair.

The first part is a hierarchical latent space. The autoencoder orders channels by prefix: first come the channels that carry the image's structure, then those responsible for detail. The beginning of the channel vector describes the overall picture, and the rest refines it.

The second part is asynchronous denoising by channel group. A single shared denoiser generates different groups at different local timesteps. The number of tokens does not grow, and the architecture does not get more complicated with a separate model for each group. According to the authors, this scheme lets the model handle structure and detail differently instead of dumping the whole latent into a single task.

Numbers from the paper

• Benchmark: ImageNet-512.
• Compression regimes: 64x and 128x.
• Comparison with existing methods: APEX is better at every tested channel count.
• Comparison with the 256-token model: APEX is better at every tested budget, with equal training cost and guided sampling.
• Trend: the lower the compute budget, the bigger the gain.
• Additional tokens: not required.

The point about budget matters more than the rest. If the gap widens precisely where compute is scarce, the method is of most interest to those who need savings first and foremost.

What to do with this in practice

For teams building their own image generation models, APEX offers a ready reason to recalculate the trade-off between token count and quality. The code is open, so the ImageNet-512 result can be reproduced without reverse-engineering the paper.

Caution is also warranted. The abstract reports results only on ImageNet-512. It does not show whether the approach works for text-to-image on large datasets, for video, or with other denoiser architectures. The cost of the autoencoder with hierarchical channels is not stated either: how much it costs to train and how sensitive it is to settings.

It is worth reading the full text and answering three questions. How exactly the schedule of asynchronous timesteps is designed. What real-world inference speedup 128x compression delivers. Whether the advantage holds as the model scales. These answers will determine whether APEX becomes a working tool or remains a neat piece of academic work.