Changelog¶
Changes to the dataset, derived files, and Python package.
2026-11-04¶
Precision-mapping-only subjects: sub-04 and sub-08¶
Added sub-04 and sub-08, who took part in precision mapping only.
They did not take part in the LAION image-viewing experiment: they have no
main-experiment sessions, no single-trial betas, no res-1pt8 files, no
object localizer, and none of the ROI sets defined by the image-viewing
data (laion, laionstim, object). sub-04’s anatomy is
ses-ConnectomAnat instead of ses-PrismaAnat.
Warning
If your code uses get_subjects, upgrade the LAION-fMRI package to
0.2.0:
python -m pip install --upgrade "laion-fmri @ git+https://github.com/ViCCo-Group/LAION-fMRI.git@main"
Versions before 0.2.0 list subjects from the bucket, so get_subjects
now returns sub-04 and sub-08 among the main-experiment subjects.
Code that loops over that list then requests main-experiment files these
subjects do not have, and fails with a DataNotDownloadedError. From
0.2.0, get_subjects() returns the main experiment’s five subjects and
get_subjects("mapping") adds sub-04 and sub-08.
Retinotopy timeseries and stimulus apertures¶
Added everything needed to fit pRF models to the retinotopy data:
the preprocessed runs of every precision-mapping session under
derivatives/preproc-mapping/, including the percent-signal-change, high-pass filtereddesc-pscHptimeseries the released pRF maps were fitted on, onres-1pt5andfsnative(see fMRI Data)per run, the design matrix as fitted (
desc-apertures_design.h5) and the frame table with its timing (desc-frames_design.tsv)the aperture libraries the designs were built from, under
stimuli/retinotopy/.
See Released stimulus files for the file layout and how to rebuild or vary a design matrix. The raw BIDS data of the precision-mapping sessions are added as well.
2026-10-19¶
Retinotopy (pRF) maps¶
Added population receptive field estimates (see Retinotopy) and
res-1pt5 masks and segmentations for every published ROI set.
Retinotopic ROIs constrained by LAION image dimensions¶
Added the laionstim ROI set: every retinotopic ROI restricted to the
vertices whose pRF centre lay inside the LAION image, on the surface and
on both volume grids, with its segmentation. See ROIs.
2026-10-15¶
Tissue masks and 1.5 mm anatomicals¶
Added cortical grey matter, subcortical grey matter, brainstem,
cerebellar cortex, white matter and ventricular CSF masks to the
anatomical derivatives, at full resolution, res-1pt8 and
res-1pt5. Added the T1w, T2w and
brain mask on the res-1pt5 grid of the volumetric pRF and
localizer maps. See Anatomical Data.
2026-10-06¶
ROI corrections¶
SPAreplacesSPLas the fourth place-selective area. This is an intended redraw, not a rename or a correction:SPAcovers more posterior vertices from theSPLit replaces, so results for the two are not comparable across versions. Theseg-placesegmentations are updated accordingly.Withdrew sub-03’s
pSTS-wordsdue to being drawn too anteriorly compared to the other subjects’. Updated itsseg-charactersegmentation accordingly.Removed sub-07’s
aLOC,dLOCandpLOCmasks, left over from an earlier object vocabulary. No segmentation referenced them.
The previous SPL and pSTS-words masks and the segmentations they
appeared in are kept under archive/rois/v1/ on the bucket.
2026-09-28¶
Corrected per-subject random splits¶
Updated the subject-specific random_0–random_4 splits to keep
identified duplicate images in the same fold, preserving split names
and fold sizes. Update the Python package to use the corrected splits.
The previous files
are archived for reproducing earlier results.
Across-subject train/test splits¶
Added pool="pooled" with tau and five cluster splits for consistent
image assignments across subjects. See
across-subject splits for usage and construction.
Corrected stimulus embedding transparency¶
Updated the CLIP, DINOv2, PEcore, and SigLIP2 embedding files for 116
out-of-distribution stimuli with non-opaque pixels. Their original
embeddings discarded transparency. The corrected embeddings composite
the images onto the experiment’s middle-grey background, RGB
(128, 128, 128), before model-specific resizing and cropping.
The remaining 24,936 embedding rows in each file are unchanged, as are the image IDs, row order, storage dtypes, and file layout. The stimulus images themselves have not changed.
The corrected files are available at the existing download URLs. To refresh an existing download, close any open embedding handles and run:
import laion_fmri
laion_fmri.download_embeddings("all")
Then reload the embeddings. This also works with older package versions: the corrected files differ in size from the originals, so the downloader replaces the cached files. No package upgrade is required for this data correction.
For reproducibility, the original files are preserved in the archive:
See Stimulus Derivatives for the embedding models and file layout.