When does the fog lift? A switching rule fixed on synthetic worlds, tested by pre-registration on pretrained image diffusion models

A diffusion sampler starts in a "fog" of noise so thick that the data are visible only through their mean and covariance. There, a closed-form Gaussian denoiser D_G built from those two moments is almost exact, and the expensive network is not needed. We ask whether a sampler can find the moment the fog lifts while it runs and without looking at the answer, and hand over to the network only from then on. The signal is the disagreement u = median |D_G − D_net|²/d between the Gaussian and the network, measured on 5% of the batch: it is low in the fog and rises when the fog lifts. Two rules read this signal: U_min switches when u exceeds 2.38 times its running minimum, and the flicker-robust U_env asks the rise to persist over a window. All their constants were fixed on synthetic worlds only, before the rules met any real network. We then carried the rules to pretrained EDM networks under pre-registered criteria (quality: FID within 3% of the full sampler M0; cost: at most 0.85 of its network evaluations; and comparison with the best analytical teleport time chosen with FID on a grid). On CIFAR-10 the transfer partly failed: four of eight pre-registered hypotheses were not supported, all on cost, and on the VE network a network-error spike in the fog made U_min switch far too early. On AFHQv2 64×64, with two networks (VP and VE) whose signal curves nobody had inspected, all 8 pre-registered hypotheses were supported, and a pre-registered replication on two further seed sets of the same networks (three seed sets, 15,000 images per method) supported all 9 further hypotheses, including on the two seed sets that nobody had seen. Both rules kept FID within the tolerance while saving about 25% of network evaluations (24.8–26.7%); in wall-clock time on an NVIDIA L4 the saving was 18–22% with the implementation used for the runs and 24–25% with an equivalent implementation that computes the Gaussian phase only on the 5% subset (measured on one batch per network). On the VE network, a finer teleport grid confirmed a pre-registered prediction to the grid point; against that best time chosen with the answer, the price of not looking was 1.2–1.6% of M0's cost (on the VP network, against a coarser grid, 7.7–8.1%, inside the registered allowance of 10%). We report the limits with the same weight: one dataset and one step count; a small but consistent quality cost on the VE network (FID +0.06, a pooled margin of only 0.013 to the tolerance); a relative FID tolerance that tightens as the number of images grows; flicker spikes that came within 0.1 of the U_min threshold; and the earlier failures of both rules (U_env on a synthetic world, U_min on CIFAR-10 VE). All code, pre-registrations with their hashes, records, and a CPU verification that recomputes every decision from the records are public: https://github.com/neginyan/fog-lift-switch

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24301908
Primary Topic
Generative Adversarial Networks and Image Synthesis
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

When does the fog lift? A switching rule fixed on synthetic worlds, tested by pre-registration on pretrained image diffusion models

Taishi Namba
1 links
Zenodo (CERN European Organization for Nuclear Research)
Generative Adversarial Networks and Image Synthesis
preprint

When does the fog lift? A switching rule fixed on synthetic worlds, tested by pre-registration on pretrained image diffusion models

Taishi Namba
preprint en

Abstract

A diffusion sampler starts in a "fog" of noise so thick that the data are visible only through their mean and covariance. There, a closed-form Gaussian denoiser D_G built from those two moments is almost exact, and the expensive network is not needed. We ask whether a sampler can find the moment the fog lifts while it runs and without looking at the answer, and hand over to the network only from then on. The signal is the disagreement u = median |D_G − D_net|²/d between the Gaussian and the network, measured on 5% of the batch: it is low in the fog and rises when the fog lifts. Two rules read this signal: U_min switches when u exceeds 2.38 times its running minimum, and the flicker-robust U_env asks the rise to persist over a window. All their constants were fixed on synthetic worlds only, before the rules met any real network. We then carried the rules to pretrained EDM networks under pre-registered criteria (quality: FID within 3% of the full sampler M0; cost: at most 0.85 of its network evaluations; and comparison with the best analytical teleport time chosen with FID on a grid). On CIFAR-10 the transfer partly failed: four of eight pre-registered hypotheses were not supported, all on cost, and on the VE network a network-error spike in the fog made U_min switch far too early. On AFHQv2 64×64, with two networks (VP and VE) whose signal curves nobody had inspected, all 8 pre-registered hypotheses were supported, and a pre-registered replication on two further seed sets of the same networks (three seed sets, 15,000 images per method) supported all 9 further hypotheses, including on the two seed sets that nobody had seen. Both rules kept FID within the tolerance while saving about 25% of network evaluations (24.8–26.7%); in wall-clock time on an NVIDIA L4 the saving was 18–22% with the implementation used for the runs and 24–25% with an equivalent implementation that computes the Gaussian phase only on the 5% subset (measured on one batch per network). On the VE network, a finer teleport grid confirmed a pre-registered prediction to the grid point; against that best time chosen with the answer, the price of not looking was 1.2–1.6% of M0's cost (on the VP network, against a coarser grid, 7.7–8.1%, inside the registered allowance of 10%). We report the limits with the same weight: one dataset and one step count; a small but consistent quality cost on the VE network (FID +0.06, a pooled margin of only 0.013 to the tolerance); a relative FID tolerance that tightens as the number of images grows; flicker spikes that came within 0.1 of the U_min threshold; and the earlier failures of both rules (U_env on a synthetic world, U_min on CIFAR-10 VE). All code, pre-registrations with their hashes, records, and a CPU verification that recomputes every decision from the records are public: https://github.com/neginyan/fog-lift-switch

Zenodo (CERN European Organization for Nuclear Research)
Generative Adversarial Networks and Image Synthesis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon