2. Agent-Assisted and Manually created LoT Image Generation
GPT6 Astra or a user manually draws an importance map, which we convert into a Level-of-Token layout for image generation. These demos show how spatial detail allocation can be designed through agent-assisted or manual control.
Generation prompt
Custom detail map → Level of Tokens
Paint the desired detail by hand, or let an agent construct the map from a text description. Black, orange, yellow, and white encode detail scores of 0, 0.3, 0.6, and 1. The canvas starts yellow. Colors describe the requested detail, not the colors of the generated image.
The scores are area-averaged onto the finest token lattice, optionally smoothed, and scaled by the overall-detail control. Starting with 8×8 blocks, the converter recursively splits a block when its maximum detail score exceeds the corresponding threshold, producing 4×4, 2×2, or 1×1 tokens. Higher scores request finer tokens; low-detail areas retain coarser tokens. The model uses the resulting token layout and text prompt, not the painted map as an appearance condition.