Original title: ToPO: Token-Conditioned Preference Routing for Attention-Based Latent Diffusion Models
Juntao Xu, Shihong Li, Hoi Fan Au, Ning Zhu · September 2026
Image-generating AI models often learn from human feedback using pairs of images where one is preferred over the other. Existing methods like Diffusion-DPO apply this overall preference broadly across every location and step in the image creation process. To improve this, researchers created ToPO, or Token-Oriented Preference Optimization. ToPO figures out specific spatial and timing paths using error contrasts from a frozen reference model. It uses text prompt tokens from the preferred image to shape spatial focus. It also adds an extra ordering step at the pixel midpoint without needing local labels or a separate reward model. In tests, ToPO beat Diffusion-DPO across all five metrics for SD-1.5, three key metrics for SDXL, and won more blind user comparisons.
Imagine cooking a stew where you add spices at different times to balance flavors. If you add all spices at once at the beginning, the stew might become overseasoned in some parts and bland in others. Here, generating an image is like cooking the stew, and ToPO is like knowing exactly when and where to add each spice to get the best taste without extra tasting or recipes.
This matters for people using AI image generators, as it makes outputs align better with human tastes without requiring detailed local labels or separate reward models, simplifying the training process.
The results are scoped to equal-update U-Net protocols and not an equal-compute comparison, so performance under different training conditions or on other model architectures is not established.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
ToPO: Token-Conditioned Preference Routing for Attention-Based Latent Diffusion Models — Juntao Xu, Shihong Li, Hoi Fan Au et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.