Effect of longer first-stage training at high compression

Determine whether extending the training duration of the PoolDINO first-stage learned pooling operator and RGB decoder improves performance at higher token-compression rates.

Background

PoolDINO learns a local affine pooling operator jointly with an RGB decoder while keeping the pretrained DINOv3 vision encoder frozen. The paper evaluates several spatial compression rates, including 4× and 16× token compression, but keeps the pooling-and-decoder first-stage training budget fixed while exploring extended training primarily for the diffusion generator. The authors therefore leave unresolved whether allocating more training to the first-stage tokenizer and decoder can improve performance, particularly at more aggressive compression rates.

References

Although we explore extended generator training, our pooling-and-decoder training budget remains fixed; whether longer first-stage training can improve performance at higher compression rates remains open.

— Pooling Representation Autoencoders for Efficient Diffusion  (2610.09242 - Calvo-González et al., 7 Oct 2026) in Conclusion, paragraph “Limitations”