Autoregressive image generators get attribute repair in the middle of the token sequence: Janus-Pro's GenEval++ accuracy rises from 24.6% to 46.8% with frozen weights
Synopsis
Autoregressive image generators often get several attributes wrong in one prompt, and adding a fixed direction to the prefix latents only reinforces an attribute the prompt already requests, rarely overriding a conflicting one. LOCUS first measures where along the image-token sequence each attribute can still be causally changed: count and position commit in early, narrow windows, while color, size and material commit in later, broader windows. It then routes test-time latent optimization to the window of the failed attribute, raising Janus-Pro's GenEval++ accuracy from 24.6% to 46.8% with the generator frozen.
Interpretation
Attribute control is located by decode-step windows in the generation sequence rather than by network layer: the same intervention flips an attribute differently depending on which window it covers. Earlier test-time control methods act at one location whichever attribute failed, and a prefix direction mainly amplifies an attribute the prompt already requests, with limited override of a conflicting one. On Janus-Pro, adding a color direction to the prefix latents overrides a conflicting prompt in only 26.7% of trials, on par with a random direction, whereas the same direction across the full trajectory overrides it in nearly all trials.
LOCUS has a verifier identify which attribute failed, then restricts test-time latent optimization to that attribute's support window. Prefix optimization helps structural attributes but misses appearance attributes, because its window covers only the former's support. With the generator, optimizer, reward, verifier and token budget held fixed, support-routed windows score 23.6 percentage points higher than late windows in GenEval++ macro accuracy; re-scoring the same 1,320 images with Qwen3.5-9B preserves the ordering.
Perspective
The conclusions cover test-time image-token latent interventions under a fixed model, reward and budget; they do not cover weight-level edits or arbitrary budgets. Cross-model evidence covers commit order only, and the band-resolved and two-dimensional support analyses are measured on Janus-Pro alone.
Several mechanism experiments use a single layer and modest sample sizes; appearance and structure are judged by CLIP and detectors respectively, a mismatch mitigated but not fully removed by a unified vision-language model check; content gating uses a fixed threshold whose sensitivity across thresholds is not characterized; and the support profiles are currently measured offline for a fixed set of attributes, with per-prompt estimation left to future work.
