Compositional generation and prompt adherence

GenEval (compositional generation) and DPG-Bench (complex-prompt adherence), higher is better. All models trained on GPIC for one epoch at 256×256.

GenEval ↑

DPG-Bench ↑

Same prompt, every model

Samples generated from identical GPIC test prompts. Switch the caption style to compare short tags with long, detailed descriptions.

Late-fusion baselinesHalcyon (ours) · 0.1B → 4B