Generating High‑Resolution Images with Only 64 Tokens: How MacTok Overcomes Posterior Collapse

In visual generation, improving inference efficiency hinges on compressing images into fewer tokens. Image tokenizers fall into two camps: discrete (e.g., VQ‑VAE, VQ‑GAN) and continuous (e.g., KL‑VAE). Continuous tokenizers retain a smooth latent space and support gradient optimization, but when aggressively compressed (e.g., using only 64 tokens) they suffer severe posterior collapse caused by strong KL regularization, leading to loss of detail in reconstructed images.

MacTok: Masked Augmenting 1D Continuous Tokenizer

The authors propose MacTok, which stems from the insight that the latent space must learn truly "semantically rich" representations to eliminate posterior collapse. MacTok adopts a dual‑pipeline design that simultaneously operates on the image space and the representation space.

Image‑Space Design

Random Masking : provides a basic reconstruction challenge, encouraging the model to build robust mappings.

Semantic Masking : evaluates the semantic importance of each image region and preferentially masks the most informative areas, forcing the model to infer high‑semantic content rather than merely filling in texture.

The optimal image‑mask ratio is found to be around 70%.

Representation‑Space Design

MacTok introduces global and local alignment . Latent tokens are aligned locally with fine‑grained patch features and globally with a condensed CLS feature, directly constraining the semantic structure of the latent space.

Experimental Results

On ImageNet 256×256, MacTok with 64 tokens achieves a gFID of 1.44, and with 128 tokens a gFID of 1.52, compressing the token sequence by a factor of 64 compared to traditional methods. Compared to SoftVQ‑VAE, the 64‑token version reduces gFID by 2.21; the 128‑token version matches the baseline REPA (gFID ≈ 1.42).

On ImageNet 512×512, both the 64‑token and 128‑token variants maintain a gFID of 1.52 and an Inception Score above 306, demonstrating that performance does not degrade despite the larger resolution.

Visualization

UMAP dimensionality reduction visualizations show that MacTok preserves distinct semantic clusters in latent space, whereas prior methods exhibit severe overlap due to collapse.

Conclusion

By integrating semantic masking in the image domain with global‑local alignment in the representation domain, MacTok creates a robust learning environment that prevents posterior collapse without altering the underlying tokenizer architecture. This dual‑space strategy offers a practical reference for future high‑efficiency, high‑quality image generation models.