Image captioning under the same pretraining budget
The 10% of T2I examples that normally have their caption dropped for classifier-free guidance get the caption moved after the image instead.
Caption dropout
caption[SOV]image tokens[EOV][EOS]
Image → text
[SOV]image tokens[EOV]caption[EOS]
T2I quality is unchanged
Caption–image alignment ↑
Qwen3-VL-Reranker-8B score, image as query, mean over GPIC Test-50K.
Captions generated by Halcyon-4B
The same model that generates images, captioning held-out GPIC test images. Pick an example below.
Generated by Halcyon-4B
Ground-truth GPIC caption