A Tilted Seesaw: Revisiting Autoencoder Trade-off for Controllable Diffusion

Pu Cao, Yiyang Ma, Feng Zhou, Xuedan Yin, Qing Song, Lu Yang · 2026

Read A Tilted Seesaw: Revisiting Autoencoder Trade-off for Controllable Diffusion by Pu Cao, Yiyang Ma, Feng Zhou, Xuedan Yin, Qing Song, Lu Yang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-21633.txt.

Description

Pu Cao, Yiyang Ma, Feng Zhou, Xuedan Yin, Qing Song, and Lu Yang revisit a trade-off inside latent diffusion models. The autoencoder in these models must do two jobs: rebuild images faithfully and provide a space where generation works well. Recent large-scale studies tilt hard toward generation scores.

The authors note that rebuild quality is reported less and less. Selections based on ablations often crown the best generation score even when rebuild fidelity drops. Their tilted seesaw view argues the balance has swung too far. Restoring weight to fidelity matters for editing tools, where a model must first copy an image well before changing it.

Catalog record

CREATORPu Cao, Yiyang Ma, Feng Zhou, Xuedan Yin, Qing Song, Lu Yang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-21633

Browse the full library · About the Tower of Babel