| ## Adaptive Canvas Policy |
|
|
| We do not use a fixed 1:1 canvas for all samples. Portrait composition prompts |
| often imply different spatial structures: some are naturally square portraits, |
| some need a vertical canvas to preserve full-body framing and upper/lower |
| breathing room, and some need a horizontal canvas to carry environmental |
| context, roads, coastlines, leading lines, or large negative space. |
|
|
| We therefore design a prompt-conditioned adaptive canvas policy. The policy |
| uses the input prompt and a learned policy state to choose the canvas size |
| before generation. Its keyword weights, decision thresholds, and candidate |
| ratios were optimized on the training set through an iterative evolutionary |
| search procedure. The selected canvas always normalizes the longer side to 1584 |
| pixels. |
|
|
| For reproducibility, we release the final learned policy state together with |
| the inference code. This allows reviewers to recover the same canvas selection |
| used in our submission. For unseen prompts, the implementation falls back to a |
| deterministic prompt-only rule policy. |
|
|
