lee101/zimage-latent-teleport-cache-adapter

Z-Image Latent Teleportation Cache Adapter

Experimental 7,312-parameter residual head for cache-guided Z-Image Turbo latent teleportation. It starts from the fitted scalar local-plus-cache gate and learns a spatial correction from:

1. the scalar-gated base movement;
2. the weighted mean of eight motion-pruned cached residuals; and
3. per-pixel/channel disagreement across those neighbours.

The output projection is zero-initialized, so an untrained head exactly matches the scalar cache gate.

Implementation, training scripts, and the paper are in lee101/cutedsl.

Held-out results

The checkpoint was trained on 160 trajectories and evaluated on a disjoint 40-trajectory fold across valid scheduler cells and horizons 1, 2, 4, and 8.

Method	relL2
Calibrated local momentum	0.2320
Scalar pruned-cache gate	0.2250
Learned cache adapter	0.2225

That is a 4.09% reduction versus local momentum and 1.11% beyond the scalar cache gate.

For six held-out step-6 to step-14 endpoint decodes:

Method	PSNR	SSIM
Local momentum	21.48 dB	0.7731
Scalar pruned-cache gate	21.84 dB	0.7881
Learned cache adapter	21.92 dB	0.7877

The 7,312-parameter head takes 3.28 ms per 64x64 latent on the recorded one-thread resident CPU benchmark. Training took 132.4 seconds on CPU. An RTX 5090 run was attempted, but the safety gate deferred it because production workloads left less than 2.5 GiB allocatable VRAM; the checkpoint therefore does not claim a GPU training time.

Files

• cache_adapter.safetensors: adapter parameters.
• config.json: architecture and retrieval protocol.
• schedule_coefficients.json: momentum and scalar-gate coefficients by step and horizon.
• metrics.json: training history, horizon breakdown, timing, and decoded-image summary.

Scope

This is an adapter for the CuteDSL latent-teleportation implementation, not a standalone image model. It is fitted specifically to Z-Image Turbo at 512px with the recorded 16-step schedule. A new model, solver, resolution, or schedule requires new trajectories and refitting. The decoded comparison is an offline endpoint forecast; a fully branched sampler benchmark remains separate.