flyvis: the connectome-constrained model of fly motion vision, what it predicts and what it has not been tested against

What is in the model

• 45,669model neurons of 64 cell types on a hexagonal lattice of 721 columns, with 1,513,231 connections (paper)
• 734free parameters learned by training; the wiring itself is fixed (paper)
• 50trained models in the authors' ensemble, all on the same connectome and task (paper; files read by us)

What it predicts, quoted from the paper

Every number comes from the Europe PMC full text of the paper (PMC11525180), read by us on 3 Oct 2026. We did not re-run the model.

A correction to our own record

The paper's words: "accurate predictions of contrast preference (FRI) were possible as long as measurements of the connection-signs were available, and that accurate predictions of the direction selectivity—but not preferred direction—could be achieved with measurements of cell connectivity, without the need for synapse count measurements".

The controls, and the one that is missing

• Random parameters Full connectome, random neuron and synapse settings: "accurate predictions of preferred contrast, but poor predictions of direction selectivity and preferred direction."
• Cell-type wiring only "task-optimized models with access to only cell-type connectivity predicted neural activity poorly."
• Merged cell types Merging all 37 excitatory, 22 inhibitory and 4 mixed cell types into three types "led to poor performance on par with the random DMN."
• Decoder alone The trained models "exhibited superior task performance to both the decoder network alone and models with random parameter configurations".
• Missing No shuffled-wiring network: no network with the same sparseness but scrambled connections was trained and compared.

"what is missing is a demonstration that the specific connectivity of the fly visual system is what enables optimal performance, rather than a generic neural network with the same level of sparseness and gross connectivity statistics of the biological network."

The controls appear only as a figure (Extended Data Fig. 2), with no number in the text, the supplement or the peer-review file and no source data, so our controls ledger records no values for them. Our search of the supplement was crude (we had no PDF tool), so it may have missed a number. In our ledger flyvis counts as a study with ablations and a decoder-only baseline but no wiring null: study row.

The 50 pretrained models

The authors publish their 50 trained models as a 3.4 MB download. We fetched it with the authors' own download script settings, checked that its sha256 equals the checksum in that script, and read each model's stored validation loss.

Licences

Try it and run it

Could we test the wiring ourselves? Our pre-check

Not now; possibly later, with a weaker question. We checked, without installing anything, whether a later run could test on a CPU whether the pretrained models need the real wiring.

• No-go under this run's memory plan. Even with the CPU-only PyTorch the install is an estimated 0.8–1.2 GB, more than our 0.5 GB scratch limit on a shared 3 GB machine.
• Flash stimuli are synthetic (drawn on the hexagonal lattice), so no movie download is needed, and the connectome file path is a setting, so a modified file can be swapped in.
• Only a limited null works without retraining. The trained synapse strengths are stored per pair of cell types, so a null must keep the same set of connected cell-type pairs, for example by shuffling the column offsets within each pair. A shuffle that changes which cell types connect, like the ones in our ledger, would need retraining on movies, which our machine cannot do.
• Conditional go for run 9 at the earliest, with its own memory plan (about 1.2 GB of scratch, nothing else running) and that retrain-free null. Such a test asks whether the trained settings still work on scrambled wiring, which is a weaker question than whether training on scrambled wiring would do as well; we will say so in the result.

Catalogue projects that use flyvis

A good grade for flyvis does not carry over: each project is graded on what it adds.

• Are fruit flies zero-shot adapters? A pretrained flyvis outputs drive same-type FlyWire neurons as inputs, column by column; scrambled wiring stops the fly steering (details).
• fly-brain (Lulzx) A a separately trained flyvis network drives the optic-lobe inputs from eye raycasts; in our browser check it passed our memory limit (details).
• flybench A optional flyvis front end for vision tasks.
• Embodied brain emulation (Eon Systems) U names flyvis as its vision; the body code is not released.

What you cannot claim from flyvis

• It is not a whole-brain model: one eye's optic-lobe motion pathways only, 64 cell types.
• It has not been shown to beat a shuffled-wiring network: the controls vary the parameters and how much of the wiring is known, not whether scrambled wiring would do as well.
• Its neurons are non-spiking and its settings are trained, so matching a recording is a prediction of a trained model, not a simulation of measured biophysics.
• The 50 models give different answers for some cell types (the paper's own cluster analysis); one model is not "the" fly.

Status

Sources

• Lappalainen et al. 2024, Nature: Europe PMC full text (XML), read 3 Oct 2026; doi.
• Peer-review file (referee quote) and the supplementary information, text-searched 3 Oct 2026.
• flyvis repository at commit 92b3845, including download_pretrained_models.py (checksum and download folder of the pretrained models).
• The authors' results_pretrained_models.zip (50 models) and license.txt, read 3 Oct 2026; third-party files were not kept.
• Our research record: the report of 3 Oct 2026, which added this page.