Experimental Design¶
The LAION-fMRI dataset uses a dense-sampling design: a small number of participants are scanned across many sessions, while seeing an extensive amount of natural images. The general logic follows NSD-style designs (few subjects, many sessions, controlled image repetitions for noise ceiling estimation), but with a broader image set based on LAION-natural (derived from LAION-5B, Roth & Hebart, 2025) that goes beyond standard scene or object datasets. Each of the 5 participants completed 30 main sessions of image viewing and one eyetracking session (ses-31), followed with a time gap of half a year by 3 additional sessions. In total, the recording took place over the course of about 1.5 years (Nov 2024 to Mar 2026) at the Max Planck Institute for Human Cognitive and Brain Sciences in Leipzig, Germany.
Experiments¶
Experiment |
Design type |
Purpose |
|---|---|---|
Main image viewing ( |
Event-related, single trials |
Single-trial responses to ~6,100 natural images per subject |
Eyetracking session (ses-31) |
Mixed (calibration + image viewing + localizer + deepmreye) |
Independent test set with concurrent eyetracking that includes
|
Todo
Do controversial stimuli need their own row in the table?
Todo
Localizer: should it be listed as a separate experiment, or only as a sub-task of ses-31?
Todo
Add retinotopy tasks to the table?
Todo
Add diffusion tasks to the table?
Main Experiment¶
In the main experiment (ses-01 to ses-30), participants viewed natural images while performing a continuous recognition task (“Have you seen this image before?”). On each trial participants pressed a button to indicate whether the image was new (first presentation) or old (already seen in a previous trial or session) with their right index finger. The task aims to provide an attentive viewing paradigm (and behavioral data on memory performance) with the main goal being to obtain reliable single-trial BOLD responses to each image.
Each subject saw around 6,204 unique images across roughly 31,856 image presentations, plus around 2,583 blank trials. The image set is split into shared and subject-unique images. The 1,492 shared images are common to all subjects: 881 of them are shown 12 times (the 12-repeat set, used for noise ceiling estimation) and 611 are shown 4 times (the 4-repeat set). The remaining 4,712 unique images per subject are shown 4 times each and are different across subjects, which maximizes the breadth of the stimulus space at the individual level.
Images were distributed across the 30 sessions following two scheduling rules. First, every image was shown at least twice within the same session at some point during the experiment, so that GLMsingle has within-session repeats to work with for HRF fitting and ridge regression tuning. No image was shown more than 3 times in the same session. Second, one repeat of the same image was always kept within ±7 sessions of its twice-per-session occurrence, so that the inter-repetition gap stays manageable for the memory task.
For details on the image content and where the images come from, see Stimulus Data. For information on which images are reserved for testing, see Train / Test Splits.
Note on incomplete repetitions¶
Most images were shown for their full number of repetitions but 8 images per participant ended up with one repetition fewer than planned. In most cases this affects subject-unique images that were only shown 3 times and more rarely, 12-repeat images that were only shown 11 times. The exact counts per subject are:
Subject |
unique (4 → 3 reps) |
shared_4rep (4 → 3 reps) |
shared_12rep (12 → 11 reps) |
|---|---|---|---|
sub-01 |
5 |
0 |
3 |
sub-03 |
4 |
1 |
3 |
sub-05 |
5 |
0 |
3 |
sub-06 |
4 |
0 |
4 |
sub-07 |
6 |
0 |
2 |
For analyses that require an exact repetition count (across-session
Noiseceiling4rep and Noiseceiling12rep maps), these images were
excluded. For analyses that average over all available repeats
(NoiseceilingAllreps), they remained. For cross-subject analyses, these
images should be excluded.
Session Structure¶
A session always consists of 12 functional runs, plus 1 structural and 4 fieldmap acquisitions. The standard layout used for all participants from ses-08 onwards was:
[S F RRRR F RRRR F RRRR F]
where S = structural (MEGRE), F = fieldmap, R = experimental run.
The structural is acquired at the start of each session, followed by a
fieldmap and then groups of 4 runs separated by additional fieldmaps.
Fieldmaps are repeated multiple times throughout the session to track B0
field drift over time.
Earlier sessions (ses-01 to ses-04) used a different layout in which the
structural was placed in the middle of the session
([F RRRR F RR S RR F RRRR F]). From ses-05 onward, the structural was
moved back to the start of the session.
Todo
Add total session duration in minutes.
Session structure overview.¶
Run Parameters¶
Duration per run |
~5 min (304 s for sessions 1-5, 309 s for sessions 6-34) |
|---|---|
Trials per run |
~94-97 (87-89 image trials + 7-8 blank trials) |
Runs per session |
12 |
Total runs (per subject) |
360 (12 runs × 30 main sessions) |
The acquisition TR is 1.9 s, but during preprocessing the data is temporally upsampled to a TR of 1.0 s for GLMsingle. So depending on whether you look at the raw or preprocessed data, the number of timepoints per run will differ (160 or 163 raw volumes vs 304 or 310 timepoints after upsampling).
Note
For MRI acquisition parameters (TR, voxel size, etc.), see MRI Acquisition.
Trial Structure¶
Each trial lasts 3 seconds. On image trials, the stimulus is shown for 2.5 s followed by a 0.5 s inter-stimulus interval. Blank trials have the same duration but no image is shown. Blank trials are distributed pseudo-randomly throughout each run and serve as a baseline condition. They also help the BOLD signal return closer to baseline between image trials which improves single-trial response estimation. A fixation cross in form of a small red dot was continuously shown also during ISI and blank trials. Participants’ responses during ISI were still counted as belonging to the previous trial image. Top button pressing indicated a new image and second button just below indicated an old image.
Trial structure for the main experiment. (placeholder — replace with actual figure)¶
Stimulus Presentation¶
Stimuli were presented through a PROpixx MRI/MEG DLP LED projector (VPixx Technologies), mirroring a BenQ display at 1920 × 1080 px and 60 Hz refresh rate onto a rear projection screen inside the scanner room. Each stimulus image was rendered at 1000 × 1000 pixels, which corresponds to about 26.5 × 26.5 cm on the projection screen, or roughly 9.2 × 9.2 degrees of visual angle. The total viewing distance was 164.8-165.7 cm (3.9-4.6 cm from mirror to eye, plus 161 cm from mirror to screen).
Software |
PsychoPy |
|---|---|
Display |
PROpixx projector + BenQ mirrored display |
Resolution |
1920 × 1080 (projector); 1000 × 1000 (stimulus) |
Refresh rate |
60 Hz |
Viewing distance |
164.8-165.7 cm |
Visual angle (stimulus) |
~9.2 × 9.2 degrees |
Scanner synchronization |
(TODO) |
Dummy scans |
0 |
Todo
Document scanner synchronization (e.g. TTL trigger, volume logging).
Localizer Experiment¶
A functional localizer (task-oloc) was acquired in the eyetracking
session (ses-31). The localizer was used to identify object-selective regions
in visual cortex (object vs scrambled) which can serve as ROI for downstream
analyses.
Todo
Specify the localizer paradigm in general, also from retinotopy sessions (classical fLOC, Stigliani et al., 2015? custom loc?).
Todo
List stimulus categories (faces, places, bodies, characters, objects, scrambled).
Todo
Describe the block design (block duration, number of blocks per category, total run duration).
Todo
Describe the task during the localizer.
Todo
Clarify which sessions the localizer was acquired in.
Task design overview.¶
Retinotopy Experiment¶
Retinotopy data is available for all LAION-fMRI subjects and was acquired in a separate study before the main 30-session image-viewing experiment started. See Retinotopy for the paradigm details and resulting maps.
Todo
Describe the retinotopy paradigm here as well, or just defer to the Retinotopy page (would duplicate content)?
Todo
Add citation to the retinotopy paper once it’s available (Satzger et al., in prep).
Behavioral Data¶
Behavioral data from the main experiment consists of the button responses
given during the recognition task, including the response (old or new) and
the reaction time per trial. These are stored in the raw BIDS events.tsv
files alongside each functional run.
An exemplary glimpse of sub-01_ses-06_task-images_run-01_events.tsv:
onset |
duration |
trial_number |
trial_type |
response |
response_correct |
response_changed |
response_time |
stim_number |
stim_name |
stim_occurrence |
|---|---|---|---|---|---|---|---|---|---|---|
12.049 |
2.9998 |
1 |
new |
1 |
1.0 |
0.0 |
0.5499 |
1 |
|
1 |
15.049 |
3.0000 |
2 |
old |
1 |
0.0 |
0.0 |
0.7000 |
2 |
|
3 |
18.049 |
3.0000 |
3 |
new |
1 |
1.0 |
0.0 |
0.9502 |
3 |
|
1 |
21.049 |
2.9999 |
4 |
new |
1 |
1.0 |
0.0 |
0.8668 |
4 |
|
1 |
24.049 |
2.9999 |
5 |
old |
2 |
1.0 |
0.0 |
0.9667 |
5 |
|
2 |
27.049 |
2.9999 |
6 |
blank |
n/a |
n/a |
n/a |
n/a |
n/a |
|
n/a |
30.048 |
2.9999 |
7 |
new |
1 |
1.0 |
0.0 |
0.8500 |
6 |
|
1 |
33.048 |
2.9999 |
8 |
old |
1 |
0.0 |
0.0 |
0.7834 |
7 |
|
2 |
36.048 |
2.9999 |
9 |
old |
1 |
0.0 |
0.0 |
1.0834 |
8 |
|
3 |
39.048 |
3.0005 |
10 |
old |
2 |
1.0 |
0.0 |
1.3001 |
9 |
|
6 |
42.049 |
2.9994 |
11 |
old |
2 |
1.0 |
0.0 |
0.9668 |
10 |
|
4 |
The full column set in the events file includes: onset, duration,
trial_number, trial_type, response, response_correct,
response_changed, response_time, stim_number, stim_name,
stim_occurrence, stim_occurrence_run, stim_occurrence_session,
stim_duration, isi_onset, isi_duration, pulse_number,
pulse_onset.
Todo
Specify questionnaire data (not shipped with release though).
Todo
Include summary statistics here? (mean accuracy, etc.)
Todo
Cross-reference more file locations.