What TRS does to text-to-image generation

Same Stable Diffusion 1.5 model, same prompt — the only change is the source noise TRS searches for. Drag to compare the baseline sample against the TRS-aligned one.

baseline TRS Baseline · SD 1.5 TRS · ours
⇄
prompt:
★ better prompt alignment & reward under the same budget

Drag the divider · click a thumbnail to switch example