Lotus-2 Geometry FP16: Depth + Normal (float16 + median degrid)

FP16 variant of Lotus-2 Geometry. Runs the FLUX transformer in float16 (instead of bfloat16): the finer mantissa (10 bits vs 7 bits) eliminates the ~1160-distinct-value depth staircase observed with bfloat16. A task-specific per-channel float median filter (scipy.ndimage.median_filter, ksize=15 for depth / ksize=11 for normal) is applied to the raw prediction to remove the FLUX-VAE 16-px patch grid.

Predicts BOTH a depth map and a normal map from a single image in one GPU call (two independent task adapters on a shared FLUX backbone — the normal map is NOT derived from depth). Returns 8-bit previews plus downloadable raw float32 .npy and 16-bit PNG for each.

512 2048
Examples
Image Process resolution (detail cap). Output detail = min(input long side, this). Higher = finer but much more VRAM/time; >1536 may OOM on this GPU. Feed an input at least this large (upscale first) to benefit.