Seeing Without Seeing: A Methodological Review of Visual Mental Imagery Reconstruction from Brain Activity

Introduction

Motivation

For two decades, decoding visual content from brain activity has mostly meant reconstructing what a person sees. This review addresses what a person imagines without an external stimulus. Imagery re-enacts perceptual representations top-down, so its signal is coarser and weaker than in perception, and its overlap with perception is strong for semantic content but weak for retinotopic detail. Methods carried over from perceptual reconstruction therefore do not transfer directly, and the content they aim to recover is never observed.

Scope and Relation to Prior Reviews

We review methods that decode or reconstruct voluntary visual mental imagery from human fMRI, EEG and MEG, admitting intracranial recordings where they bear on the same question (§ 2.2 and § 2.5). Electrophysiology shows imagined content to be decodable, most robustly from parieto-occipital alpha-band EEG activity [1, 2], reinstated in single neurons [3] and usable for closed-loop control [4], with null results also reported [5]; our search identified no image-level reconstruction from it scored against an image-level target (§ 6.4). Seen-image reconstruction has been reviewed [6, 7], and its pipelines and corpora [8, 9] enter here only as an upper bound and as reusable components.

The closest prior review [6] raises zero-shot and spurious-prediction concerns for presented stimuli, where the scoring target exists by construction. Without such a target, further elements are needed: transfer across a change of state (§ 5.3), an admissibility criterion (§ 2.2), the proxy gap (Eq. (16)) and checkable reporting standards (§ 7.4); Table 1 compares the other reviews, including the survey of Guo et al. [7]†. From Shirakawa et al. [10] we adopt the results on output dimension collapse, identification inflation and the recovery check and ask which carry over when no ground-truth image exists.

A Unified Pipeline and Taxonomy

Almost every method instantiates a two-stage map in which a decoder $h_{\psi}$ transforms the neural response $\mathbf{x}$ into a representation $z$ and a generator $G$ renders it into a reconstruction $\hat{I}$ (Eq. (3), Figure 1). Methods differ along two design axes (§ 4.1) and fall into four families (§ 4.2).

Main Challenges

Two issues run through every family: perception-to-imagery transfer and target-free evaluation (§ 5.3 and § 7.1). They meet in the problem of spurious reconstruction (Figure 2): a frozen, high-capacity generator can synthesise a photorealistic image weakly coupled to $\mathbf{x}$ while feature, semantic

Vector (PDF)

and realism scores are satisfied largely by the prior [10], a concern a reanalysis has raised for a published imagery result [16]⁴. § 7.3 treats the mechanism, and Eq. (17) gives a statistic that separates read-out from prior.

Contributions

• Imagery reconstruction as the object. Earlier reviews cover the neuroscience of imagery [11, 12, 13, 14] or seen-image reconstruction [7, 6]; here internally generated vision is treated as a decoding problem in its own right.

• A taxonomy and an admissibility criterion. Four families with bounded composability, and a criterion crossing working-memory separability with memory source ($L1/L2/L3$).

• Evaluation without ground truth. A synthesis of published imagery performance (Table 5) and, as an author proposal, a normalised neural dependence that consolidates existing null-signal controls [10, 17] (Eq. (17)), a diagnostic quantity not yet validated on independent data.

Table 5: Reported performance on paired vision / imagery benchmarks, all from fMRI perception-transfer decoders. Rows compare within a panel only (note e); values are verbatim except V−I and $\rho$ (notes a, f).

Panel	Method	Vision overall	Vision simple	Vision complex	Imagery overall	Imagery simple	Imagery complex	Imagery conceptual	V−I	$\rho$	Source
Panel A — NSD-Imagery; human 2AFC, $n = 500$ raters; trained on subj. 1, 2, 5, 7; chance 50%; % overall (simple / complex / conceptual)	MIRAGE	84.22	80.77	87.66	78.30	73.93	83.19	77.68	+5.9	0.83	[17] Tab. 2
Panel A — NSD-Imagery; human 2AFC, $n = 500$ raters; trained on subj. 1, 2, 5, 7; chance 50%; % overall (simple / complex / conceptual)	MindEye1	84.29	79.32	89.32	73.00	71.01	82.28	65.68	+11.3	0.67	[28] Tab. 2
Panel A — NSD-Imagery; human 2AFC, $n = 500$ raters; trained on subj. 1, 2, 5, 7; chance 50%; % overall (simple / complex / conceptual)	MindEye2	83.05	75.54	90.56	56.96	50.21	64.83	55.74	+26.1	0.21	[28] Tab. 2
Panel A — NSD-Imagery; human 2AFC, $n = 500$ raters; trained on subj. 1, 2, 5, 7; chance 50%; % overall (simple / complex / conceptual)	Brain-Diffuser	80.13	71.79	88.46	73.95	68.20	82.70	71.01	+6.2	0.80	[28] Tab. 2
Panel A — NSD-Imagery; human 2AFC, $n = 500$ raters; trained on subj. 1, 2, 5, 7; chance 50%; % overall (simple / complex / conceptual)	iCNN	74.52	75.85	73.19	66.15	66.81	70.04	61.60	+8.4	0.66	[28] Tab. 2
Panel A — NSD-Imagery; human 2AFC, $n = 500$ raters; trained on subj. 1, 2, 5, 7; chance 50%; % overall (simple / complex / conceptual)	Takagi & Nishimoto	49.58$^{c}$	48.95	50.20	52.89$^{c}$	57.26	48.31	53.14	−3.3	n.d.	[28] Tab. 2
Panel B — NSD-Imagery; automated metrics on simple + complex stimuli; PixCorr (higher better) and CLIP two-way %, chance 50%	MIRAGE	0.221	66.69	0.104	57.46	+9.2	0.47 / 0.45	[17] Tab. 1
Panel B — NSD-Imagery; automated metrics on simple + complex stimuli; PixCorr (higher better) and CLIP two-way %, chance 50%	MindEye1	0.218	65.34	0.086	54.72	+10.6	0.39 / 0.31	[17] Tab. 1
Panel B — NSD-Imagery; automated metrics on simple + complex stimuli; PixCorr (higher better) and CLIP two-way %, chance 50%	MindEye2	0.161	65.93	0.036	50.78	+15.2	0.22 / 0.05	[17] Tab. 1
Panel B — NSD-Imagery; automated metrics on simple + complex stimuli; PixCorr (higher better) and CLIP two-way %, chance 50%	Brain-Diffuser	0.107	58.31	0.064	54.07	+4.2	0.60 / 0.49	[17] Tab. 1
Panel B — NSD-Imagery; automated metrics on simple + complex stimuli; PixCorr (higher better) and CLIP two-way %, chance 50%	iCNN	0.224	49.03	0.108	41.72	+7.3	0.48 / n.d.	[17] Tab. 1
Panel B — NSD-Imagery; automated metrics on simple + complex stimuli; PixCorr (higher better) and CLIP two-way %, chance 50%	Takagi & Nishimoto	−0.013	43.01	−0.006	40.08	+2.9	n.d.	[28] Tab. 1
Panel C — Deeprecon; pairwise identification, each pipeline scored by its own similarity$^{e}$; conv2–fc6$^{d}$; 3 subjects; chance 50%; % ± SD	Koide-Majima	90.7 ± 13.5	89.6 ± 8.9	75.6 ± 22.3	72.9 ± 23.2	+15.1	0.63	[37]
Panel C — Deeprecon; pairwise identification, each pipeline scored by its own similarity$^{e}$; conv2–fc6$^{d}$; 3 subjects; chance 50%; % ± SD	iCNN, re-run	64.3 ± 30.6	66.7 ± 32.9	50.4 ± 32.2	51.4 ± 31.8	+13.9	0.03	[37]
Panel D — Deeprecon, iCNN as reported in [29]; 3 subjects; chance 50%; identification % by pixel correlation / human judgement ($n = 13$ raters)	Natural, with DGN	76.1	97.0	qualitative only	—	n.d.	[29] Figs. 3B, 8C
Panel D — Deeprecon, iCNN as reported in [29]; 3 subjects; chance 50%; identification % by pixel correlation / human judgement ($n = 13$ raters)	without DGN	79.7	96.0	—	—	—	[29] Figs. 6C, 8D
Panel D — Deeprecon, iCNN as reported in [29]; 3 subjects; chance 50%; identification % by pixel correlation / human judgement ($n = 13$ raters)	Artificial shapes	70.5	91.0	51.9	83.2	+18.6 / +7.8	0.09 / 0.81	[29] Figs. 6C, 8D

• A candidate reporting checklist. Ten proposed reporting items with illustrative pass criteria (Table 6), an author proposal rather than an established standard, applied in an exploratory manner to ten studies to check whether the items are reported in a checkable form, not to rank study quality (Table 7), followed by open problems.

Proposed item	Compute	Illustrative pass criterion	Failure it closes
S1 Null-signal baseline	$\Lambda$, $\tilde{\Lambda}$ (Eq. (17)) under test-time mismatch: re-run the deployed pipeline on another trial’s $x$ with generator, samples and seeds fixed; between-trial and within-category nulls for generative and retrieval pipelines; label permutation reported separately	$\Lambda_{\mathrm{within}} > 0$ and $\Lambda_{\mathrm{between}} > 0$, each with an episode-level permutation test, $p < 0.05$ over $B$ stated permutations ($B \ge1000$ where feasible); for label outputs, above-chance held-out accuracy under the same test; never $E$ alone	Prior-driven synthesis scored as read-out (Eq. (23))
S2 Zero-shot categories	Evaluate on categories, and where possible exemplars, held out from decoder training	Above chance on held-out categories, reported separately	Classification mistaken for generalisation [10, 31]
S3 Dual-stream metrics	Low-level (PixCorr, SSIM [95]) and semantic (CLIP) scores computed separately	Both reported, not merged into a composite nor selected post hoc	Dissociation hidden by one scalar (Eq. (22))
S4 Vividness covariate	Trait (VVIQ) and, where available, per-trial vividness; score regressed on it	Distribution and score–vividness relation reported, even when null	Proxy gap confounded with method quality (Eq. (16))
S5 Generator recovery	Feed true latents of the perceptual stimuli into the frozen generator and score recovery with the reconstruction metrics	Recovery exceeds the reported reconstruction score, or the generator bound is stated	Generator loss attributed to the decoder [10]
S6 Train–test structure	Semantic clusters of the training targets and train–test overlap; $r_{95}$, $\kappa_{r95}$ and the effective-rank ratio $r_{\mathrm{eff}}(\hat{Z}_{\mathrm{te}})/r_{\mathrm{eff}}(Z_{\mathrm{te}})$ (Eq. (20))	Reported; if overlap is substantial or the ratio well below 1, S2 also on a cluster-disjoint split	Output dimension collapse (Eq. (19))
S7 No best-of-$N$ selection	Fix samples per trial and the selection rule before evaluation	Rule stated; qualitative examples drawn at random; seed variability reported	Selective reporting [16]
S8 Episode-level validation	Split and permute by imagery episode, run or session; augment after splitting	No episode on both sides of a split; cross-block or cross-session result for block designs	Autocorrelation leakage [83, 42, 45]
S9 Admissibility statement	Trial-level separability crossed with memory source (L1/L2/L3), Sec. 2.2	Both stated, with the trial timeline that justifies them	Maintenance or recall reported as generation
S10 Eye state	Eye state in imagery and in perception; eye tracking or EOG regression where feasible	Stated for both; any difference addressed for the alpha band	Alpha carrying ocular state, not content

Study	Family	S1	S2	S3	S4	S5	S6	S7	S8	S9	S10
Naselaris et al. 2015 [30]	Ia	(✓)	✓	(✓)	✗	—	✗	✓	✓	✓	—
Senden et al. 2019 [36]	Ia	✗	(✓)	(✓)	✗	—	✗	(✓)	✓	✓	—
Shen et al. 2019 [29]	Ib	✗	✓	(✓)	(✓)	✗	✗	(✓)	✓	✓	—
Koide-Majima et al. 2024 [37]	Ib	✗	✓	(✓)	✗	✗	✗	✗	✓	✓	—
Reddy et al. 2010 [39]	II	(✓)	✗	—	✗	—	✗	—	✓	✓	✓
Horikawa & Kamitani 2017 [31]	II	(✓)	✓	—	(✓)	—	✗	—	✓	✓	✓
Xie et al. 2020 [1]	II	✓	✗	—	✗	—	✗	—	✓	✓	(✓)
Kwon et al. 2025 [42]	II	✗	✗	—	✗	—	✗	—	✗	✗	✗
Kneeland et al. 2025 [28]	III	(✓)	(✓)	(✓)	✗	✗	✗	✓	✓	✓	—
Kneeland et al. 2026 [17]	III	(✓)	(✓)	(✓)	✗	✗	✗	✓	✓	✓	—

The first three contributions and the performance comparison synthesise the literature; the neural dependence and the reporting checklist are proposals of the authors.

Per-cell evidence for the tables, the electrophysiological results, the derivation of Eq. (21) and further figures are given in the appendices following the references.

Scope, Terminology, and Problem Formulation

Scope

This review covers methods that decode or reconstruct internally generated visual content, as distinct from reconstruction of perceived images [18, 19]. Most decoders therefore face a perception-to-imagery transfer problem, and because the imagined content is latent, an evaluation-without-ground-truth problem. Seen-image reconstruction, such as MindEye [20], MindEye2 [21], latent-diffusion decoders [22, 23], MinD-Vis [24] and mesh-guided diffusion [25]†, is treated as a perceptual baseline whose numbers enter only as the paired vision condition of an imagery benchmark.

Terminology: Defining the In-Scope State

_Voluntary visual mental imagery._ The core object is the deliberate, top-down generation of a quasi-perceptual visual experience without the corresponding external stimulus, under intentional control, typically initiated by a label, a word or a previously seen exemplar, with retained awareness that the experience is imagined [13, 11]. Neighbouring states are separated in Table 2, whose “Scope” column is implemented at screening by criteria E2 and E3 (§ 2.5).

State	Volition	Ext. stim.	Pseudo-GT	Aware $\ne$ perc.	Scope
Voluntary imagery	Vol.	None	Yes (cue)	Yes	Core
Working memory	Vol.	None (removed)	Yes (seen)	Yes	Boundary
Dream imagery	Invol.	None	No	Usu. no	Boundary
Hallucination (clin.)	Invol.	None	No	Often no	Out
Attempted/covert	Vol.	None	Variable	Yes	Out (mention)
Perception (contrast)	—	Present	Yes (GT)	—	Out (contrast)

_Imagery, working memory and hallucination._ Imagery and working memory are both voluntary and stimulus-independent at measurement and share representational format in visual cortex [26], but working memory maintains an item just perceived, which supplies a de facto ground truth, whereas imagery can generate content specified only by a symbolic cue; delay-period decoding [27] is therefore a boundary case. Hallucination is involuntary, usually lacks experimenter-specified content and is often believed real, so it is excluded.

A two-dimensional admissibility criterion. Whether an exemplar is shown inside the imagery trial is a necessary but not sufficient test: a paradigm may show no exemplar in the trial yet ask the participant to retrieve a picture seen an hour earlier, which differs from generating content never presented as an image. We therefore cross it with the memory source of the imagined content:

L1 — maintenance. An exemplar presented within the trial is held across a delay [27].

L2 — recall of a learned exemplar. A specific image memorised before the session is recalled on a symbolic cue, as in NSD-Imagery [28], Deeprecon [29] and the memorised artworks of [30]; the seen picture is the image-level scoring target.

L3 — generation. Content is specified by a category, concept or description [31, 2, 28], so no image-level target is presented.

Within L3 we distinguish purely conceptual generation, for which no image-level target exists even in principle, from compositionally constructed targets, for which a ground-truth image can be generated by rule from the cue; the compositional cueing of § 8.1 belongs to the latter.

The two dimensions are written together $(\checkmark\cdot\mathrm{L2}, \times\cdot\mathrm{L1}, \ldots)$ in Table 3 and Table 4. The criterion, for example, classifies the largest public EEG imagery dataset as entangled, since the target picture is shown inside every imagery trial [32]. Because purely conceptual L3 content has no image-level target, image-based scoring is possible only under L1, L2 or compositionally constructed L3 targets, and the last are not used in the reviewed corpus. Where an exemplar is shown inside the trial, the memory source is L1 unless the instruction directs imagery away from it (Figure 3); which cells are empty by definition and which are merely unoccupied is discussed in Appendix § 16.§ 1.

Typical modality	Applicable source states	Supervision source	Primary output target	Anchor studies	Admissibility
Ia. Encoding-model inversion into pixel space	fMRI only; no EEG/MEG/iEEG anchor as of Sept. 2026	Percep.-trained, inverted (cross-state)	Vol. imagery (scenes, shapes)	Thirion [35]; Naselaris [30]; Senden [36]	✓ · L2 [30, 36]
Ib. Encoding-model inversion through a feature space	fMRI only	Percep.-trained, inverted with a deep-feature or generative prior	Vol. imagery; retrieved faces	Shen [29]; Koide-Majima [37]; Lee & Kuhl [38]	✓ · L2 [29, 37, 38]
II. Discriminative decoding and representational geometry	fMRI · EEG · MEG · iEEG (only family anchored in every modality)	Imagery-trained or cross-state (labels)	ID / classif. / retrieval; RSA	fMRI: [39, 31, 40], [41]†. EEG: [1, 2, 42]; null: [5]. MEG: [43]. iEEG: [3]. WM boundary: [27]	✓ · L3 [31, 2], [41] ✓ · L2 [39, 40, 1] ✓ · L2c [3] ✗ · L1 [5, 42, 27, 43]
III. Generative cross-decoding	fMRI; one exploratory EEG attempt, not image-scored (Sec. 6.4)	Percep.-transfer / hybrid	Vol. imagery (transfer from perception)	Pixel + semantic recon. (diffusion / AR)	NSD-Imagery [28]; MIRAGE [17]; Spera [44]†; EEG, exploratory: [45]	✓ · L2 [28, 17, 44] ✓ · L3d [28] ✗ · L1 [45]
IV. Alignment and adaptation (meta-level)	fMRI (whole-brain), cross-subj.; EEG only via foundation models (outlook)	Imagery ↔ perception alignment; cross-subj. adaptation; foundation-model transfer (outlook)	Vol. imagery; low-SNR regimes	Meta: improves the mapping feeding the targets above	Hyperalignment [46, 47]; MindAligner [48]; NeuroSwift [49]; code conversion [50]; Stable-Mind [51]†; Spera [44]†; outlook: [52, 53]†, [54]	—b ✓ · L2 [44]

Dataset	Modality	Subjects	Imagery task and cue	Target status	Admissibility
Imagery-focused, fMRI
NSD-Imagery [28]	fMRI 7T	8	6 simple shapes, 6 complex scenes (5 from shared1000), 6 concept words; memorised letter cues; 432 trials per subject	Proxy c; image target except concepts	✓ · L2 simple, complex; ✓ · L3 conceptual
GOD, Horikawa & Kamitani [31]	fMRI 3T	5	Category-name cue, eyes closed; arbitrary exemplars of the category	Proxy c; seen-image GT for training	✓ · L3
Deeprecon [29]	fMRI 3T	3c	Word cues for 25 images memorised beforehand (10 natural, 15 artificial)	Proxy c; seen-image GT for training	✓ · L2
Imagery-focused, EEG / MEG / iEEG
VI-BCI EEG [32]	EEG, 32 ch	22 (2 sess.)	10 pictures; 17 s trial: picture 4 s, mask, imagery 4 s; VVIQ	Proxy Ic; class label only	✗ · L1
Action-imagery EEG [42]	EEG, 64 ch	15	4 upper-limb actions; 17 s trial: video 5 s, rest 5 s, imagery 5 s; E2 boundary case; not openly released	Proxy Ic; class label only	✗ · L1
Semantic-concept EEG [85]	EEG, 124 ch	12	3 concepts (pictorial, orthographic, auditory), each perceived, then imagined in the trial; VVIQ, BAIS-V	Proxy Ic; class label only	✗ · L1
EEG + fNIRS imagery [86]	EEG, 64 ch; fNIRS (+7 EEG)	12	36 animals or tools; 0.6 s picture and mask, then imagery of one’s own exemplar; animal / tool labels	Proxy Ic; class label only	✗ · L3
MEG / iEEG imagery	MEG; iEEG	—	None openly released as of Sept. 2026; none outside fMRI pairs imagery with image-level targets	—	—
Perceptual corpora used as cross-decoding substrate
NSD [8]	fMRI 7T	8	Natural scenes, up to 30,000 trials per subject	Perceptual GT	—
BOLD5000 [87]	fMRI 3T	4	∼5,000 natural images	Perceptual GT	—
THINGS [88, 89, 9]	fMRI; MEG; EEG	3; 4; 50, 10	THINGS object images; EEG by rapid serial visual presentation	Perceptual GT	—
EEG-ImageNet∗ [84]	EEG	6	Block-design viewing of 40 ImageNet classes	Perceptual GT	—

Problem Formulation

Let $\mathbf{x}$ be the response on a single imagery trial, a voxel pattern for fMRI or a channel-by-time array for EEG/MEG, often reduced to band-limited power because the content-bearing imagery signal is oscillatory rather than time-locked [1, 2], and let $\mathbf{s}$ be the imagined content. No external image $I = s$ is presented: the experimenter observes only a cue $c$, which may be a label, a word or a previously seen exemplar $I_c$. The cue is therefore a proxy target and $s$ is never observed; an image-level proxy exists only under L1 and L2.

A decoder estimates a class, a ranking of candidates in an embedding space, or a reconstruction in pixel space or a semantic latent such as CLIP [33]. Supervision is (a) imagery-trained on scarce, low-SNR imagery pairs; (b) perception-transfer, fitting a decoder or a forward encoding model $g_{\phi}: I \mapsto\mathbf{x}$ on perception and applying it to imagery, which relies on representational overlap [14]; or (c) hybrid. Every method with imagery-reconstruction numbers in Table 5 falls in (b). Encoding-model inversion learns $g_{\phi}$ on perception and recovers content under a natural-image prior [34]:

A Unified Taxonomy

The primary axis is the method family (§ 4.2). Six cross-cutting dimensions provide a shared vocabulary for characterising each family (Table 3): modality (fMRI, EEG, MEG, iEEG); source state (voluntary imagery, working-memory maintenance, dreaming); memory source (L1, L2, L3); supervision source (imagery-trained, perception-transfer, hybrid); output target (classification, retrieval, pixel-level or semantic reconstruction); and stimulus type (simple shapes and letters, natural scenes, concept words).

Review Methodology

Type of review and search. This is a criterion-guided methodological review, that is, a narrative review with explicit criteria, rather than a PRISMA systematic review: no protocol was pre-registered and no two independent raters screened. Instead, we state explicit criteria, give a uniform admissibility verdict for every tabulated study and tie each tabulated cell to a source sentence (Appendix A). We searched Web of Science, Scopus, PubMed, IEEE Xplore, DBLP, arXiv and bioRxiv, with Google Scholar for forward citations, using four query groups: imagery decoding (A), generative reconstruction (B), evaluation methodology (C) and ethics (D). A–C supply the tabulated studies, and § 6.4 re-runs A restricted to EEG, MEG, iEEG, ECoG and sEEG. Searches closed on 30 September 2026; the corpus effectively starts with inverse retinotopy in 2006 [35], earlier work entering as background [11, 3, 55]. Only English full texts were considered.

Inclusion and exclusion criteria. A record is included if it I1 decodes or reconstructs internally generated visual content from human brain activity measured without the corresponding stimulus; I2 reports a quantitative read-out (an accuracy, a reconstruction scored by an operator of Eq. (4), a representational-geometry statistic, or, for datasets, a technical validation); I3 records humans with fMRI, EEG, MEG, iEEG or single units; I4 is peer-reviewed or a preprint meeting P1–P3; and I5 has an English full text. It is excluded if it E1 concerns perceived stimuli only; E2 concerns non-visual content such as imagined speech or motor imagery (visualised actions instructed as visual are admitted as boundary cases); E3 concerns involuntary or pathological states; E4 is non-human or simulation-only; E5 appears in a non-archival venue; or E6 is a methods paper without an imagery application.

Preprint policy. A preprint is cited only if no peer-reviewed version existed at 30 September 2026 (P1) and anchors a claim only if no peer-reviewed source supports it (P2); every preprint anchor is flagged †, and no standard or conclusion rests on a preprint alone (P3).

Limitations and disclosure. Selection was by the first author, with no independent second screener, so the corpus is a traceable selection whose error has not been estimated, and its criteria were fixed during the search. Two cited perceptual methods [49, 25]^† share two authors with this review; they enter under the same criteria as other work, as a Family IV anchor and a perceptual baseline respectively.

Background and Cognitive Foundations

Neural Substrates: Visual Mental Imagery versus Perception

A shared but asymmetric substrate. Imagery re-enacts perceptual representations across the ventral visual hierarchy [56]. Classifiers trained on seen stimuli generalise to imagined ones most robustly for high-level features [39, 31], shape-specific codes reinstated during imagery localise to anterior lateral occipital complex [40], and imagery and perception share representations of content and location [57]. Early visual cortex nonetheless carries depictive imagery information: encoding models fit on perception identify which of thousands of candidate images a participant recalls, even from V1 or V2 voxels [30]. In EEG, shared object representations are carried by parieto-occipital alpha [1], and medial-temporal neurons responsive to a picture fire again when it is imagined [3].

The overlap is neither complete nor uniform. Imagery is generated from frontoparietal and higher visual areas projecting back onto sensory cortex [14, 11, 56], and generative-feedback accounts explain why seen and mental images can elicit distinct codes [58]. Early-visual reactivation is coarser and weaker than in perception, and the share of the perceptual representation recovered by imagery grows along the hierarchy [31, 14]. Imagery representations also emerge later, are more sustained and are less precise than perceptual ones [43, 2].

Computational consequence: the transfer problem. Because semantic representations overlap strongly and retinotopic detail weakly, decoders transfer most readily at the semantic level [12], which is why modern pipelines condition on high-level embeddings. To compare transfer across metrics we use the retention ratio

the fraction of above-chance vision performance that survives the move to imagery, with $\mathcal{E}_{\mathrm{chance}} = 50\%$ for two-alternative protocols and 0 for pixel correlation. On Table 5 (Figure 4), MIRAGE retains $\rho= 0.83$ of its human-rated margin but 0.47 of its pixel correlation, and Brain-Diffuser 0.80 against 0.60, whereas iCNN retains 0.48 of its pixel correlation while its CLIP score is below chance in both conditions (§ 7.2). $\rho$ is unstable when its denominator is small, and intervals can be formed only where dispersion is reported (Panels C and D); since human ratings are not purely semantic, the contrast between streams should not be generalised beyond these methods.

Individual Differences and Boundary Conditions

Vividness and aphantasia. Imagery vividness, usually measured by the VVIQ [55], tracks the neural overlap between imagery and perception both trial-wise and across individuals [59]. At the extreme lies aphantasia, the congenital inability to generate voluntary imagery [60, 61]. Chang et al. [62] decode imagined content from V1 in aphantasic individuals as well as in controls, yet the patterns do not cross-decode with perception (26.2%, $p = 0.274$, against 43.9%, $p < 0.001$, in controls); sound-driven decoding in early visual cortex [63] is diminished in aphantasia [64]; decodable V1 representations can occur without subjective imagery [65]; and 7 T fMRI shows reduced fusiform–frontoparietal connectivity in aphantasia [66]. Decodability and phenomenology can therefore dissociate, so imagery ability should be reported as a covariate (S4); no study in Table 5 relates its scores to it.

Reality monitoring and the working-memory boundary. Candidate signals that mark an experience as internal are evaluated jointly by a higher-order circuit [67], and the dominant cue appears to be signal strength, since sufficiently vivid imagery is misattributed to perception [68]. Content may thus be strong enough for a classifier yet too weak for experience, and a decoder is not bound by the reality-monitoring threshold (§ 9.1). Working-memory contents are likewise decodable from V1–V4 [27] in representations partly shared with imagery [26], which is why every tabulated study carries both components of the admissibility verdict.

A Unified Taxonomy Framework

The Mind’s-Eye Decoding Pipeline

A common two-stage map. Decoding maps $\mathbf{x}$ to a target through a model-side decoded latent $z$ in a representation space $\mathcal{Z}$, such as a CLIP or CNN-feature embedding [69]:

where $h_\psi$ is a neural-to-latent decoder and $G$ a generator. With $G=\operatorname{id}$ the latent is used directly for classification or retrieval, which is where the electrophysiological literature currently sits (§ 6.4); in the inverted direction a forward encoding model $g_\phi:\mathcal{I}\to\mathcal{X}$ is learned instead.

Two design axes. Methods differ along two axes that the family labels of Table 3 partly conflate (Fig. F.7). The direction of inference is either encode-and-invert, fitting $g_\phi$ on perception and searching image space to match the response, or decode, mapping the response to a latent with $h_\psi$. The generator is none ($G=\operatorname{id}$), per-trial iterative optimisation, or a frozen, amortised pretrained model applied in one pass. Family Ia is encode-and-invert with iterative search; Family Ib is decode with iterative generation, since its neural stage is a feature decoder (Eq. (8)) and what it inverts is the image-to-feature network; Family II is decode without a generator; Family III is decode with a frozen amortised generator; and Family IV, a transform on the input, is orthogonal to both axes. Ib thus shares its neural stage with Families II and III, which is why its exposure to prior-driven failure resembles III more than Ia (§ 7.3).

Evaluation under a missing target. Because $s$ is unobserved, every method is scored against the proxy, the cue $c$ or the exemplar $I_c$, by an operator

whose target-free character, absent in perception where $I_c=s$, is analysed in Section 7.

The Four Method Families

Family I — Iterative inversion. In Family Ia a forward encoding model $g_{\phi}$ is fit on perception, mapping image features $\varphi(I)$ to the response, and imagined content is recovered by inverting it under a prior:

where $R$ is an image prior and Eq. (6) is the MAP counterpart of Eq. (1). With a fixed, low-level feature map such as Gabor wavelets or a retinotopic basis, the inversion is a search over candidates or a direct pixel reconstruction,

as in [35, 30, 36]. In Family Ib the brain is decoded into the hierarchical features of a deep network, and pixels or a generator latent are optimised to match them:

as in [29, 37, 38]. Both variants optimise per trial and are cross-state, since imagery enters only at inference.

Family II — Discriminative decoding and representational geometry. Here $G=\mathrm{id}$ and $h_{\psi}$ predicts a label or regresses onto a fixed embedding $z_{\mathrm{target}}$ for retrieval:

Representational similarity analysis (RSA) is a model-free variant that tests a second-order isomorphism between neural and feature geometries [70, 57].

Family III — Generative cross-decoding. A neural-to-latent map is trained on perception or hybrid data to reach the conditioning space of a pretrained generator built on denoising diffusion [71] or latent diffusion [72], and imagery is cross-decoded through the frozen generator:

with $\mathcal{L}_{\mathrm{CLIP}}$ a contrastive term. The imagery step is a distribution shift, so seen-image performance does not guarantee imagined-image performance [28, 17], and strong priors make plausible but spurious reconstructions most likely here [10].

Family IV — Alignment and adaptation (meta-level). Family IV learns a transform $T$ that aligns representations across subjects, sessions, sites or states, so that Families I–III remain usable with scarce imagery data, either by functional alignment,

or by aligning imagery activity into a frozen conditioning space,

together with cross-subject adaptation and, as an outlook, neural foundation models (§ 5.4). It defines no decoding target of its own and pre-composes with any family.

Method Families in Detail

Family I: Iterative Inversion

Estimator. Reading the data term of Eq. (6) as a Gaussian likelihood and $R(I)$ as a negative log-prior exposes Ia as maximum-a-posteriori inference:

No imagery labels enter training; the burden rests on the fidelity of the forward model and the tractability of the inverse problem.

Representative work. Inverse retinotopy set the template by inverting a perception-fit forward model [35]. Voxel-wise encoding with Gabor and semantic tuning then showed that low-level features of remembered scenes are present in imagery: the encoding model identified which of five memorised artworks a participant recalled among random photographs well above chance [30, 33]. The same logic recovers imagined letters [36]. In Family Ib, iCNN optimises pixels so that their features in a deep backbone such as VGG [73] match brain-decoded features, optionally under a deep generative prior [29, 74, 75]; imagined artificial shapes gave rudimentary but recognisable reconstructions, whereas imagined natural images were reported as not well reconstructed [29]. A Bayesian reformulation with a VQGAN prior and decoded CLIP and VGG features [37] raises pairwise identification of imagined natural images from 50.4% to 75.6%, with two qualifications (Table 5, notes d and e): conv1 is excluded, and each pipeline is scored with its own weighted similarity, a circularity raised by a reanalysis [16]⁴.

Trade-offs. Ia recovers only what a retinotopic model encodes, so a high score requires genuine low-level information, but imagery’s early-visual signal is coarse and weak (§ 3.1), and the need for a spatially resolved forward model confines the family to fMRI. Ib gains expressivity from a learned feature decoder and a prior and is correspondingly exposed to the prior–evidence tension of § 7.3. Both are non-convex, so results depend on initialisation, prior weight and restarts, and the choice of which run to display is addressed by standard S7.

Family II: Discriminative Decoding and Representational Geometry Estimators

The direct mode optimises Eq. (9); the model-free mode correlates representational dissimilarity matrices (RDMs),

Representative work. Reddy et al. [39] trained classifiers on four object categories and tested every combination of perception and imagery: within-perception 67%, within-imagery 50% and cross-decoding 47% and 52%, against chance at 25%, with only perception-to-perception above chance in retinotopic cortex. Horikawa and Kamitani [31] removed the closed-set restriction by decoding hierarchical DNN features, identifying categories not used in training. Their imagery condition is L3 by instruction, but participants were first shown exemplar images of each category, and the imagined categories are those of the perceptual test session, so recall of seen exemplars cannot be excluded. The same approach separates imagery from perception of real objects [76], recovers sleep-onset and dream content [77, 78], and relates decodable patterns to reported image qualities [41]†. In the reviewed corpus this is the only family anchored in every modality reviewed, with EEG [1, 2], MEG [43] and single-unit anchors [3, 4] and a published EEG null [5] (§ 6.4).

Kwon et al. [42] decode four imagined actions from EEG (72.3% against 25% chance); because participants were instructed to imagine without muscle engagement, we admit the study under E2 as a boundary case (§ 6.4).

Trade-offs. Because semantic representations overlap strongly across states and classification needs only category-level separability, Family II is data-efficient and comparatively robust to low SNR. Its ceiling has the same origin: the output is a label or an embedding, retrieval is bounded by the candidate set, RSA yields no read-out, and identification depends on the distractor set (§ 7.2). This is the output type that EEG and MEG currently support.

Family III: Generative Cross-Decoding

Mechanics and anchors. Family III maps the response to the conditioning code of a frozen generator, typically a CLIP image or text embedding [72, 69, 79, 80, 81], so imagery decoders inherit a high-capacity prior and reconstruct in one pass. NSD-Imagery evaluates NSD-trained decoders [23, 20, 21, 29, 22] on paired imagery and finds seen-image performance largely decoupled from imagined-image performance, with simple linear or multimodal decoders generalising better than complex ones [28]. MIRAGE ridge-regresses into multimodal text-plus-image embeddings that condition a frozen cascaded diffusion model and improves imagery reconstruction under human forced choice [17]. Outside fMRI there is one exploratory EEG attempt [45].

The nature of the shift. The map $h_{\psi}$ is fit on perception and deployed on imagery, and whether that change is attenuation or a change of code determines which remedies apply. Write the response in state $r \in\{\mathrm{perc}, \mathrm{img}\}$ as

with $A^{(r)}$ a region-specific gain. Under the shared-code hypothesis $H_0$, $f^{\mathrm{img}} = f^{\mathrm{perc}}$ and imagery differs only in gain and noise; under $H_1$ the code itself changes. Even under $H_0$ with $A^{\mathrm{img}} < A^{\mathrm{perc}}$ the Bayes-optimal $\mathbb{E}[z \mid x]$ differs through shrinkage, so a change in $p(z \mid x)$ alone does not separate the two cases; and because the latent available on imagery trials is $z(I_c)$ rather than $z(s)$, the apparent imagery conditional also moves with the proxy bias of § 7.1.

The neural evidence is consistent with $H_1$ in early visual cortex and compatible with $H_0$ plus gain elsewhere: the overlap is attenuated in early visual cortex [31, 14]; perception-trained classifiers transfer in object-responsive but not in retinotopic cortex [39]; imagery aligns with late rather than full perceptual processing [43, 1, 2]; and aphantasic imagery is decodable in V1 yet does not cross-decode with perception [62], which gain alone explains less naturally than an altered code.

Published benchmark values add a diagnostic signal (Figure 4). They are compiled across two reports [28, 17] and compared under a common human forced-choice framing; this is a cross-study published-value comparison, not a unified benchmark experiment. In these values, MindEye2 falls from 75.54% to 50.21%, near the 50% chance level, on simple shapes, and from the highest complex-scene score in the compilation, 90.56%, to 64.83% [28]. Across the five methods that differ from chance the drop spans 5.9 to 26.1 points and the ordering of methods changes, while MIRAGE scores within 0.1 points of the best vision value and loses 5.9 points on imagery [17]. The change in ordering is consistent with $H_1$ but also with attenuation acting on heterogeneous estimators, so it does not discriminate between them; that vision rankings do not carry over to imagery is a testable hypothesis suggested by these values rather than an established result. A direct test is possible with existing data (§ 8.2).

Spurious reconstruction. The general mechanism is analysed in § 7.3. The risk is highest here because a frozen, high-capacity generator can satisfy feature, semantic and realism scores from little neural information [10]. MIRAGE addressed this by drawing human-experiment distractors from decoded random brain patterns, so that target and distractor carry the prior equally [17]; this approximates the between-trial null of Eq. (17), which is why Table 7 credits the study partially on S1. Since the risk grows as the signal weakens, null-signal controls would be a precondition for extending the family to EEG or MEG.

Family IV: Alignment and Adaptation

Composability and its limits. Let $M$ be the base mapping of any family (the inversion operator of Family I, $h_\psi$ of Family II or $G \circ h_\psi$ of Family III). Alignment leaves the family’s loss untouched and substitutes $x \mapsto T(x)$, giving $M \circ T$:

(with Family III),

(with Family I).

The composition is always well formed, because $T$ is fitted separately. Whether it preserves what $M$ does depends on $T$ mapping the target domain into the region on which $M$ was fitted, which presupposes a shared latent geometry. Across subjects within one state this is well supported [46, 47]; across states it is what H$_1$ (§ 5.3) denies, and alignment into a perception-derived space can import a perceptual bias. If $T$ projects onto the subspace that $h_{\psi}$ reads most confidently, composition can raise proxy scores while lowering the neural dependence of Eq. (17), so imagery-specific alignment should be reported with the null-signal control applied after composition.

Representative work. Hyperalignment learns a common space by Procrustes-style transforms [46, 47]; learned alignment [48], lightweight subject adaptation [49], code conversion without shared stimuli [50] and source-free or unseen-subject decoding [82, 51] extend it to limited data. The imagery-specific instance of Eq. (12) aligns imagery activity into a frozen diffusion space [44]†, while neural foundation models [52, 53, 54]† have so far been evaluated on general tasks rather than imagery. No published imagery score isolates an alignment layer against the same pipeline without it, so the family has no entry in Table 5.

Datasets, Paradigms, and Benchmarks

Experimental Paradigms

Cued imagery dominates: a label $c$ or a previously seen exemplar $I_c$ specifies the content on each trial [31, 29]. Free imagery, accessible only through retrospective report, is closer to real-world use but hard to decode or score, and working-memory delay paradigms [27] are the L1 paradigms of § 2.2. When the cue is an image just viewed, a perceptual trace of $I_c$, a maintenance signal and the generated imagery are entangled in $x$, and a decoder scored against $I_c$ cannot separate read-out from carry-over. This is most pronounced outside fMRI: the largest public EEG imagery dataset presents the target picture for 4 s before each 4 s imagery period [32], and the action-imagery study a 5 s video before imagery [42]; both designs aim to remove the afterimage, not the maintenance signal.

Modality-Specific Characteristics

fMRI offers the retinotopic access that underwrites pixel- and feature-level reconstruction, but limited per-trial SNR, which is why every benchmark in Table 5 averages repetitions; 7 T fMRI improves SNR enough to reveal phenotype differences [66]. EEG and MEG provide millisecond resolution suited to the late, sustained dynamics of imagery [43], but volume conduction and artefacts make image-level reconstruction much harder. Intracranial recording combines millisecond and cellular resolution but samples medial temporal cortex in clinical populations and has yielded no public imagery corpus [3, 4]. Design matters as much as modality: in a block design slow drifts are confounded with class labels, and a reanalysis [83] of an early 40-class EEG dataset [84] showed accuracy falling to chance under a rapid-event design.

Principal Datasets and Reported Performance

Purpose-built imagery datasets are small in trials and stimuli rather than in participants: NSD-Imagery has 432 vision and imagery trials over 18 stimuli per subject [28], against up to 30,000 perception trials in NSD [8]. The three fMRI imagery datasets cue learned material symbolically and are separable at the trial level, whereas all four EEG datasets of Table 4 are entangled and supply only class labels, so no electrophysiological analogue of NSD-Imagery was identified.

What image-based scoring can and cannot test. Apart from compositionally constructed L3 targets, which no study in the reviewed corpus uses, image-based scoring is available only under L1 or L2, and the reviewed corpus occupies only L2 (Figure 3): wherever an imagery reconstruction is scored against an image, as for Deeprecon’s memorised pictures [29] and their reanalysis [37], NSD-Imagery’s simple and complex stimuli [28, 17, 44], the memorised artworks of [30] and the letters of [36], the participant recalls an exemplar shown before the session. In the reviewed corpus, no L1 paradigm has been carried to image-scored reconstruction, and no study has built a scorable image target for L3 content, although L3 is decoded at the classification level [31, 2] and NSD-Imagery’s conceptual reconstructions are human-rated against words [28, 17]. The gap is thus that generation has not been put to an image-level test, rather than that it has failed one. This matters because retrieving a stored exemplar can, in the limit, be satisfied by an a Trial level ✓ separable, ✗ entangled, ? unclear, — none crossed with memory source (L1 / L2 / L3); evidence in Table A.8.

b Vision + imagery; 144 further attention-run trials are released but excluded by the authors.

c Five in later public sessions of the same laboratory [50].

d A picture is shown in the trial, but participants visualise their own exemplar of the concept (§ 2.2).

∗ Block design with class-contiguous trials, hence a leakage risk. expressive prior from a category label (§ 7.3), whereas generated content has no stored target to collapse onto (§ 8.1).

Internal validity of the imagery benchmarks. Five of NSD-Imagery’s six complex stimuli come from the NSD shared1000 set, which participants viewed repeatedly [28]; this is not leakage, since the set is held out from training, but it places the condition at the recall end of L2 in a semantically clustered corpus [10]. Deeprecon’s artificial shapes lie outside the natural-image training distribution [29] and are therefore the more diagnostic half of Panel D. Neither benchmark decodes single trials, so every figure in Table 5 is a multi-trial quantity (note g).

EEG/MEG/iEEG Imagery: What Exists and What Does Not

In the reviewed corpus, image-level reconstruction scored against image-level targets has been reported only from fMRI, while electrophysiology has focused on classification, representational analysis and BCI control (the corpus by year and modality is shown in Fig. F.10). Part of this division is principled, since Family Ia needs a spatially resolved forward model that sensor-level EEG and MEG do not provide. For semantic-level generative decoding, however, the main constraints appear to be missing image-level imagery targets and single-trial SNR: semantic reconstruction of perceived images already succeeds from MEG [90] and EEG [91], and the semantic stream is the one that transfers to imagery (§ 3.1).⁷

What has been established. Imagined content is decodable with an oscillatory carrier: EEG cross-classifies imagined and perceived objects in the parietooccipital alpha band [1], and extends to whole scenes and graded scene properties [2]. Imagery has a distinct temporal profile, its MEG representations emerging later, more sustained and less precise than perceptual ones [43]. Reinstatement is visible in single neurons, which patients can drive volitionally to control a display [3, 4], the only closed-loop imagery demonstration we found. Visual imagery has also been evaluated as an EEG-BCI control strategy with mixed outcomes [92, 93, 94], and a time-resolved EEG study found no above-chance decoding during imagery at any time point [5], consistent with the recoverable signal lying in band-limited power over a long window.

Why reconstruction is harder here. For Family I the limit is physical: a sensor-level topography is a volume-conducted mixture whose cortical source is ill-posed, and the decodable imagery signal is band-limited power [1, 2], which supports closed-set classification but cannot constrain pixels through Eq. (13). For Family III the limits are data and SNR: imagery offers no event to lock to beyond the cue and its onset jitters across trials, so averaging blurs the dynamics of interest, and jitter alone can mimic a sustained representation [14].

Ocular state and leakage. Imagery is often performed with eyes closed and perception with eyes open [39, 31], and eye closure raises posterior alpha, the band that carries decodable imagery content, so eye state should be reported (S10). Long trials, class-blocked runs and a fixed cue–imagery–rest template also expose imagery paradigms to the temporally autocorrelated leakage shown for block-designed EEG [83]; for example, one study augments trials before a 60/20/20 split [42], and another cuts each imagery period into segments before splitting [45]. Splitting and permuting by imagery episode or run addresses this (S8).

What is missing. No openly released MEG or intracranial visual-imagery dataset was identified as of 30 September 2026, and the public EEG datasets [32, 85, 86] carry class labels rather than image-level targets. The one EEG synthesis attempt [45] scores its images only through a classifier, slightly above the 2.5% chance of 40 classes, and shows the target before

date. Class counts, chance levels and validation units of all electrophysiological studies are listed in Table D.9. each imagery period (L1), so it remains exploratory. An electrophysiological dataset would enable Family III if it (a) cued material learned before the session symbolically; (b) paired each trial with its image-level target and released the stimuli; (c) recorded a within-subject perception condition; (d) logged eye state and vividness per trial; (e) randomised class order and released episode labels; and (f) added an L3 condition.

Evaluation Methodology

The Target-Free Evaluation Setting

The proxy gap. In perception $I_c = s$, so a distance between $\hat{I}$ and $I_c$ is a distance to the ground truth. In imagery the evaluator can compare $\hat{I}$ only with the proxy through the operator of Eq. (4), and for a distance-type operator $d$ the triangle inequality bounds the discrepancy by the distance between cue and content (Figure 5a):

For similarity-type operators, such as identification, CLIP scores and human ratings, no such bound holds in general, and the relevant quantity is the proxy bias $b(\hat{I}; I_c, s) = \mathcal{E}(\hat{I}, I_c) - \mathcal{E}(\hat{I}, s)$. The bias depends on the reconstruction itself and is systematic rather than zero-mean, varying with imagery ability, cueing and memory source (§ 6.1 and § 3.2), so it cannot be averaged away and confounds comparison across paradigms. The memory-source levels order $\mathbb{E}[\Delta]$ coarsely: smallest for L1, where the proxy was on screen moments earlier, larger for L2, and largest for purely conceptual L3, where the proxy is a word and the exemplar is the participant’s own.

Neural dependence. Eq. (16) prices the proxy but does not indicate whether a reconstruction used the brain at all. Consolidating the randomised-data control of [10] and the randomly decoded distractors of [17], we propose the neural dependence of a reconstruction procedure, a diagnostic quantity rather than a validated metric, defined as the drop in its score when the pairing between neural data and target is destroyed with everything else fixed, together with its normalised form (Figure 5b):

with expectations over trials and the generator’s stochasticity and the same seeds for real and null inputs. $\widetilde{\Lambda}$, the fraction of the above-chance score attributable to trial-specific signal, is intended to be comparable across operators and studies and is reported unclipped. As a quantity proposed here, $\Lambda$ has not been validated on independent data: its sensitivity, its specificity and its robustness to cue leakage and to eye state remain to be established. The null is test-time mismatch: the trained pipeline is driven with another trial’s activity (training-label permutation answers a different question and is reported separately). With $x_{\mathrm{null}}^{\mathrm{between}}$ a trial of any other stimulus and $x_{\mathrm{null}}^{\mathrm{within}}$ a trial of a different stimulus from the same category or set, a pipeline that decodes category alone achieves $\Lambda_{\mathrm{between}} > 0$, whereas only $\Lambda_{\mathrm{within}} > 0$ indicates within-category dependence. $\Lambda$ can be estimated from existing data by re-running the pipeline on null inputs; label outputs, cue leakage and cost are treated in Appendix § 16.§ 2. Reporting $(\mathcal{E},\Lambda_{\mathrm{between}},\Lambda_{\mathrm{within}},\widetilde{\Lambda})$ with an episode-level permutation test is proposed item S1; it is designed to separate the two panels of Figure 2, where a spurious reconstruction holds $\mathcal{E}$ high while $\Lambda\approx0$. Even $\Lambda_{\mathrm{within}} > 0$ is necessary rather than sufficient, since it does not certify the visual detail displayed. In the exploratory checklist application below, an S1 entry records only whether a corresponding null control is reported, not whether that control is adequate.

Output dimension collapse. If the neural-to-latent map is a ridge regression from training activity $X_{\mathrm{tr}}$ onto training latents $Y_{\mathrm{tr}}$, every prediction is a combination of the training latents:

The linear span is restrictive only when $N$ is smaller than the latent dimension; the relevant constraint is effective: shrinkage concentrates $\mathbf{m}$ on the dominant directions of the training targets, so that when those targets are clustered the predictions fall near the subspace spanned by the cluster centroids. Shirakawa et al. call this output dimension collapse and find that the CLIP text features of NSD images form roughly 40 clusters with substantial train–test overlap [10]. Two quantities measure it:

where $P_r$ projects onto the $r$ leading principal directions of the centred training latents, with $r = r_{95}$ explaining 95% of their variance, and $C_Z$ is the covariance of a set of latents $Z$. A small $r_{95}$ with $\kappa_{r_{95}}$ near 0.95 marks a test set inside a low-dimensional, clustered training distribution; a ratio $r_{\mathrm{eff}(\hat{Z}_{\mathrm{te}})}/r_{\mathrm{eff}(Z_{\mathrm{te}})}$ well below 1 marks predictions less diverse than their targets (S6). This argument applies directly to linear/ridge latent mappings; for nonlinear decoders it is a heuristic diagnostic. Where it applies, collapse compounds the proxy gap: a translator confined to a few clusters and scored against a cluster-level proxy can appear successful while recovering little that is trial-specific.

A Taxonomy of Evaluation Metrics

(1) Identification and retrieval. Identification asks whether $\hat{I}$ or $z$ is closer to the correct proxy than to distractors, with chance $1/n$ for $n$ alternatives, and its main pitfall is the distractor set. If the test set has $k$ equally sized categories and the decoder recovers category but nothing within it, so that a same-category distractor is identified at chance and a different-category one with probability $q \in[0.5,1]$, the large-$n$ pairwise accuracy is

derived with its finite-sample form in Appendix § 15 and plotted in Fig. E.6 [10]. With $q=1$ it is 75% for $k=2$ and 95% for $k=10$, so a pairwise accuracy alone is not evidence about the particular image. Protocol choice also matters: under human forced choice and an automated CLIP two-way comparison, both at a nominal 50% chance, iCNN reaches 66.15% and 41.72% on the same imagery reconstructions [28, 17]. An accuracy reported without its $n$, distractor rule and scoring space therefore cannot be ranked against another.

(2) Low-level and feature-metric alignment. Feature metrics score $\mathcal{E}$ as a distance in a fixed representation: pixel correlation and SSIM [95]; features from AlexNet [96], residual networks [97], DINO [98] or EVA-02 [99]; learned perceptual distances [100]; and CLIP scores [69]. FID [101] and the Inception Score [102] gauge realism, not fidelity to $s$. For imagery, low-level fidelity and semantics are only weakly coupled (§ 3.1),

so the two streams should be reported separately. Panel B of Table 5 shows the dissociation: iCNN has the highest pixel correlation in both conditions (0.224, 0.108) and the lowest CLIP score of the five significant methods (49.03%, 41.72%).

(3) Human rating and (4) phenomenological correspondence. Raters reward naturalness, so a fluent, prior-driven synthesis can score well while weakly tied to $x$. Drawing the distractor from a reconstruction of a different brain pattern by the same pipeline [28, 17] equalises the prior and estimates $\Lambda_{\mathrm{between}}$; a same-category distractor would estimate $\Lambda_{\mathrm{within}}$. Decodability can also be scored against experience, trial-wise or via the VVIQ [55]; a vividness correlation supports genuine read-out [59], but its absence does not disqualify [62, 65]. NSD-Imagery records per-trial vividness [28], yet no study in Table 5 reports this analysis.

Spurious Reconstruction

The mechanism. A frozen, high-capacity text-to-image prior (Eq. (10)) can synthesise a photorealistic, category-consistent image from little information in $\mathbf{x}$, so the output becomes a blend of classification into a trained category and prior-driven generation [10]. As the neural dependence falls, it collapses onto the prior’s typical sample for the inferred class (Figure 2):

which remains fluent while decoupled from $\mathbf{x}$, and which is testable because its antecedent is estimated by the null-signal control.

Why standard metrics may not detect it. Feature and CLIP scores can be met by the class-typical sample, realism ratings reward the prior’s fluency, and cross-category identification can be solved by category alone (Eq. (21)). Convincing reconstructions have been obtained even from randomised brain data by sampling repeatedly and keeping the best output [10]. A reanalysis of a published imagery result raises concerns about selective multi-run reporting and circular evaluation [16]⁴; under P3 it carries no conclusion alone, but Panel C of Table 5 does score each pipeline with its own weighted similarity [37] and is therefore a within-study contrast rather than an accuracy comparable with Panels A, B or D. As modern priors make realistic outputs common regardless of fidelity [10, 15], realism is weak evidence, which is why $\Lambda$, insensitive to realism, is reported instead.

The recovery check. Feeding the true latent features of a target image into the generator tests whether the image comes back; if it does not, the generator bounds the pipeline whatever the decoder does. Text-guided pipelines return semantically similar but perceptually different images in this check, whereas feature-matching inversion passes it [10]. For imagery only the generator-side half transfers: the check is run on the perceptual stimuli of the same corpus, with the paired vision condition as an upper bound under the same generator and subjects (S5).

The chain of failure. A clustered corpus induces collapse (Eq. (19)); collapsed latents are rendered as fluent class-typical images (Eq. (23)); categorical test structure and fluency-rewarding raters inflate their scores (Eq. (21)); and the score is then read as fidelity. Each link has a standard in Table 6 (Fig. F.8): S6 for the corpus, S2 and S5 for translator and generator, S3 and S7 for the metric, and S1 for the inference. Together they operationalise for imagery the correctives proposed by Shirakawa et al., namely diverse training data, compositional representations and evaluation designed to expose prior fill-in [10].

A Candidate Reporting Checklist

We distil the analysis into a candidate reporting checklist of ten proposed items, each stated in Table 6 as what to report, how to compute it and an illustrative pass criterion, and apply it in an exploratory manner to ten studies in Table 7. The items are proposals of the authors, not established standards.

Provenance, scoring and clarifications. S1–S4 were fixed before the exploratory application and S5–S10 added during it, so the application is not fully independent of the corpus it scores. The distractor-rule requirement of S3 below was specified after the initial scoring; S3 cells are therefore scored from published distractor descriptions rather than a complete re-check, and are entered conservatively (Table 7). A study reports all proposed items in a fully checkable form when every applicable item is ✓. Three readings apply to the illustrative pass criteria: S3 also requires the number of alternatives $n$, the distractor-sampling rule and the proportion of same-category distractors (Eq. (21)); the S5 recovery score should be reported as the pipeline’s ceiling, with the ratio of reconstruction to recovery score; and S1 should state the cost of the null re-runs and, for imagery-trained decoders with a cue on screen, include the cue-only control (Appendix § 16.§ 2).

The exploratory application. In this exploratory application, checkable reporting is more frequent where a proposed item matches existing practice, as with S2 in work building on zero-shot generic decoding [31], and less frequent where no convention requires the quantity: no study reports S5 or S6, one reports S1 in a fully checkable form and five partially, and S3 is reported at most partially under the conservative scoring described above. These observations describe reporting in ten studies, not the quality of those studies or of the field.

Reading the columns. S1 is ✓ where the null described for proposed item S1 in Table 6 is reported, which records that a corresponding control is reported, not that it is adequate: for label outputs, whose mismatch null scores chance by construction, a permutation test of held-out accuracy at the trial level of a randomised design, as in the sign-permutation cluster statistics of [1]. (✓) marks a between-trial null without the within-category one, as in the randomly decoded distractors of [17], or permutation tests against chance whose relation to the test-time mismatch of Eq. (17) the report does not make checkable. S2 is ✓ where held-out categories are evaluated by design [31, 29, 37, 30] and (✓) where test stimuli are held out yet belong to the training cluster structure [28, 17]. S3 is — where the output is a label; elsewhere it is scored from the published descriptions of metrics and distractor protocols rather than a complete re-check against the distractor-rule requirement of § 7.4, so studies reporting both metric streams are entered as (✓). S8 and S9 are ✗ for [42] (§ 5.2), and S10 is — for fMRI, ✓ where eye state is specified and monitored [39, 31], (✓) where its alpha-band implication goes unanalysed [1].

Summary. Of the two proposed items that no study reports even partially in this application, S6 requires only an analysis of the stimulus corpus and of predictions already computed, and S5 only access to the generator, so both could be supplied retrospectively from existing data.

Open Problems and Future Directions

Data Scarcity and the Generative Gap

The reconstruction values in Table 5 rest on two benchmarks and eleven participants [28, 29], only four of whom supply the decoders evaluated on NSD-Imagery, and outside fMRI no imagery dataset with image-level targets was identified (Sec. 6.4); the missing ingredient there is the target rather than the sample size.

Generation. The L3 case, content composed rather than retrieved, has no image-based evaluation (Sec. 6.3). Three routes could make it scorable. (i) Compositional cueing: cue novel combinations of familiar attributes, such as a red circle above a blue square, so that a ground-truth image can be built from the cue without being shown. (ii) Post hoc externalisation: have participants draw, select or describe their exemplar after the trial; decoding recalled video content into descriptive text [103] suggests that a verbal target is feasible. (iii) Within-participant consistency: score whether a reconstruction matches the same participant’s other instances of a repeated L3 cue better than other participants’ instances, an identification problem with a well-defined chance level and no image target. Route (i) yields compositionally constructed L3 targets in the sense of § 2.2, whereas routes (ii) and (iii) address purely conceptual content.

More perceptual data inherits the cross-state shift and does not raise the effective dimension of predictions fitted on a clustered corpus (Eq. (19)); cross-state alignment into a space where a trained decoder and a frozen generator remain valid [44]† is a promising alternative, provided the composed system is reported with S1.

The Nature of the Cross-State Shift

Section 5.3 posed the question as a shared code with state-specific gain ($\mathrm{H}_{0}$) against an altered code ($\mathrm{H}_{1}$) [14, 13]. NSD-Imagery and Deeprecon both contain matched perception and imagery trials within subjects, so the question can be addressed without new data (Fig. F.9). One would fit the perception mapping $W^{\mathrm{perc}}$ and compare, on held-out imagery trials, a gain-only model $\operatorname{diag}(\mathbf{a}) W^{\mathrm{perc}}$ with region- or voxel-wise gains against a freely refitted $W^{\mathrm{img}}$ under matched regularisation, by cross-validated prediction of the proxy latent, separately for early visual and high-level ventral cortex. A gain-only win supports $\mathrm{H}_{0}$, and a refit win beyond what its extra parameters buy supports $H_1$. Reddy et al. ran a classification-era version of this comparison [39]; we found no study repeating it for generative decoders. A theory of achievable fidelity under such a shift would need to reproduce the retention ratios of Figure 4(b), including a seen-to-imagined drop that, in the published values, differs by a factor of 4.4 across methods trained on the same corpus, and the two-stream dissociation (Eq. (22)) suggests stream-specific bounds, tighter for semantics than for pixels.

Individual Phenotypes and Fine-Grained Dynamics

Because the proxy bias depends on imagery ability, which spans a continuum from aphantasia to hyper-vivid imagery [60, 55, 59], modelling ability is a prerequisite for unbiased evaluation, yet no study in Table 5 reports scores by ability, even where per-trial labels exist [28]. 7 T fMRI [66], intracranial recordings [43, 3] and intracranial foundation models [104] offer finer access to phenotype differences and imagery’s sustained dynamics.

Real-Time Closed-Loop Use

Nearly all work is offline, whereas communication and assistive use require a closed loop. The amortised pass of Eq. (10) suits real time better than per-trial optimisation, though diffusion generation [72] still incurs latency. Every number in Table 5 is a multi-trial average, whereas a closed loop has none to average, which should lower $\Lambda$ and raise the prior’s share of the output; reporting single-trial performance would be a natural first step. The one closed-loop demonstration [4] used a four-way choice, and EEG-BCI work with visual imagery reports similarly few degrees of freedom [92, 93, 94]. Because a hallucinated reconstruction fed back to a user could mislead them, the null-signal control is here also a safety consideration.

Ethical Considerations

The literature in this section is the normative Group D of § 2.5, to which the study-level criteria do not apply.

Mental Privacy

A perceived image existed in front of the person; an imagined one need never have entered the world. Imagery decoding therefore concerns content that is private in a strong sense [105], and without an external record the person cannot contest what a decoder claims to have read [106]. Since a decoder is not bound by the reality-monitoring threshold (§ 3.2), a result may also concern content the person would sincerely deny having imagined [62, 65].

Consent and Cross-Subject Decoding

A practical safeguard against unconsented decoding has been that a decoder needs hours of the person’s own labelled data; semantic language decoding from fMRI failed when trained on other people or when the participant resisted [107]. Family IV reduces that requirement [46, 50, 82, 49]. These results are perceptual and consented, so the concern is prospective, but a method that reduces the data needed from a person also reduces what that person must agree to. Similarly, every reconstruction in Table 5 is trained on perceptual data, such as NSD [8], consented to as a perception experiment; consent documents could state whether derived models may be applied to internally generated content.

Neurorights and Regulation

The neurorights agenda holds that existing rights may be insufficient for technologies that access neural activity directly [108, 106]. Chile added to its constitution in 2021 a requirement that technological development safeguard brain activity and the information derived from it [109], applied by its Supreme Court in 2023 to a consumer EEG device [110]; UNESCO adopted a recommendation on the ethics of neurotechnology in 2025 [111]; and Colorado and California treat neural data as sensitive in their privacy statutes [112, 113]. Because mental states can also be inferred from non-neural signals, whether “neural data” is the right regulatory target is debated [114]; imagery decoding, whose target is a mental state, sits on that boundary, which suggests framing safeguards around consent and purpose [115].

Misuse of Hallucinatory Reconstructions

A reconstruction collapsed onto a class prior could be misread as a picture of what a person remembered or intended. The best results in Table 5, 78.30% for MIRAGE [17] and 83.2% for imagined shapes [29], are two-alternative, multi-trial figures from cooperative participants imagining pre-specified content; they do not support inference about uncooperative subjects, single trials or unspecified content, and on this record generative imagery reconstructions are not suitable as evidence of memory or intent [108, 115]. In the assessment of unresponsive patients, cited as motivation in the benchmark literature [28], errors can run both ways: a fluent image generated largely by the prior could be taken as evidence of awareness, and a failed reconstruction, which can have many causes unrelated to the person, as evidence of its absence.

Conclusion

Reconstructing what a person imagines differs from reconstructing what a person sees: imagery is generated top-down, is weak in early visual cortex and is never observed, so every decoder is trained and judged through a proxy. Three observations follow from the published record reviewed here. In a published-value comparison, method ordering changes between seen and imagined images, which suggests that vision performance is not a reliable proxy for imagery performance; whether the neural code itself changes can be tested with existing data. Image-based scoring is confined to maintained or recalled exemplars and compositionally constructed targets, the reviewed corpus occupies only recall, and content generated from a concept has not yet been tested at the image level. Within the reviewed corpus, image-scored reconstruction has been reported only from fMRI, a pattern consistent with the lack of image-level targets in electrophysiological imagery data.

Because a high-capacity generator can produce convincing images from little neural information, proxy fidelity alone cannot separate read-out from prior fill-in. Reporting the normalised neural dependence, a diagnostic quantity proposed here, with a within-category null, alongside the other proposed items of Table 6, is one way to address this. In this exploratory application, none of the ten studies reports all proposed items in a fully checkable form; this is not a quality ranking, and the two items that no study reports (S5 and S6) could be computed retrospectively. Because the object is private experience, progress should be accompanied by protections for mental privacy and by caution against treating generative reconstructions as evidence.

The Corpus and Its Per-Cell Evidence

Table A.8 itemises the imagery studies and datasets of Table 3, Table 4 and Table 7 of the main text with the evidence behind each admissibility verdict; Appendix § 12 and Appendix § 13 do the same for Table 1 and Table 7, and Section Appendix D collects the electrophysiological results. For GOD [31], the L3 verdict reflects the instruction to imagine arbitrary exemplars; participants had first been shown exemplars of each category, and the fifty categories are those of the perceptual test session (§ 5.2).

Evidence for Table 1

Each verdict in Table 1 rests on the presence or absence of a dedicated section in the review concerned. [11] is organised by neural substrate and contains no section on decoding method, evaluation or governance. [12] is organised by mechanism and clinical application, with decoding appearing as evidence within the mechanism sections. [13] contains a substantial treatment of objective measurement, which we score as partial on the target-free column because it concerns measuring imagery rather than evaluating reconstructions, and no section on method taxonomy, standards or ethics. [14] is organised by the dimensions of perception–imagery overlap and closes with outstanding questions rather than standards. [15] is a validity critique with a sustained treatment of what decoders do and do not establish, which we score ✓ on target-free evaluation, while whether its cautions amount to an auditable protocol is a judgement we record as partial. [7] organises fMRI-to-image work as a three-stage pipeline, a unified taxonomy, and treats perception throughout. [6] is organised by latent representation, compositionality and modularity and raises zero-shot and spurious prediction for seen images, both scored partial for the reasons in note (d) of Table 1.

Evidence for Table 7

S1. [1] tests held-out pairwise accuracy against chance with sign-permutation cluster statistics; for label outputs this is the mismatch null of Eq. (17) (Sec. Appendix F.2), hence ✓. [17] constructs human-experiment distractors by decoding randomly selected brain activity patterns, so that target and distractor are equally shaped by the prior, and separately reports reconstruction quality across guidance strengths; this estimates $\Lambda_{\mathrm{between}}$ but

Study / dataset	Modality	Family	Trial	Src	Evidence for the verdict
Thirion et al. 2006 [35]	fMRI	Ia	✓	L2	Imagery of previously presented Gabor-like patterns, cued without re-presentation
Stokes et al. 2009 [40]	fMRI	II	✓	L2	Letter cue (X or O) initiates imagery of a shape learned before the run; no exemplar in the trial
Reddy et al. 2010 [39]	fMRI	II	✓	L2	Auditory name cues; participants close their eyes before imagery blocks and open them before perception blocks, verified by eye tracking
Naselaris et al. 2015 [30]	fMRI		✓	L2	Imagery runs cue mental images of five artworks memorised in advance while fixating a blank grey screen
Lee & Kuhl 2016 [38]	fMRI	Ib	✓	L2	Faces retrieved from paired-associate learning completed before scanning
Horikawa & Kamitani 2017 [31]	fMRI	II	✓	L3	Cue is an array of 50 object names; participants imagine arbitrary exemplars of the red-marked category with eyes closed for 15 s
Shen et al. 2019 [29]	fMRI	Ib	✓	L2	Word cues for 25 images memorised before the experiment; word–image pairs viewable only during inter-run rest
Senden et al. 2019 [36]	fMRI	Ia	✓	L2	Imagery of letter shapes learned before the imagery runs
Shatek et al. 2019 [5]	EEG	II	✗	L1	Four images shown in sequence within each trial, then a numeric retro-cue names which to imagine

Study / dataset	Modality	Family	Trial	Src	Evidence for the verdict
Xie et al. 2020 [1]	EEG	II	✓	L2	Imagery cued symbolically for objects learned before the imagery blocks
Chang et al. 2025 [62]	fMRI	II	✓	L2	Lateralised imagery of learned patterns on cue; aphantasia and control groups
Stecher & Kaiser 2024 [2]	EEG	II	✓	L3	Three-sentence written scene descriptions; the matching photographs are shown only afterwards, in a separate perception task
Gao et al. 2026 [32]	EEG	—	✗	L1	Each 17 s trial: 3 s fixation, 4 s picture, mask, 4 s imagery, 4 s rest
Kwon et al. 2025 [42]	EEG	II	✗	L1	Each 17 s trial: 2 s rest, 5 s action video, 5 s rest (stated purpose: afterimage removal), 5 s imagery
Shimizu & Srinivasan 2022 [45]	EEG	III	✗	L1	500 ms image, 500 ms noise mask, then 10 s imagery, within each trial
Kneeland et al. 2025 [28]	fMRI	III	✓	L2/L3	Only a single-letter cue appears during imagery; all 18 cue–stimulus pairs memorised before scanning. Conceptual stimuli are words with no image
Kneeland et al. 2026 [17]	fMRI	III	✓	L2/L3	Same benchmark and protocol as above
Koide-Majima et al. 2024 [37]	fMRI	Ib	✓	L2	Reanalysis of the Deeprecon imagery data; cueing as in [29]
Dijkstra et al. 2018 [43]	MEG	II	✗	L1	Retro-cue task: a face and a house are shown for 0.8 s each, then a noise mask and a 0.5 s cue naming which of the two to imagine for 3.5 s
Kreiman et al. 2000 [3]	iEEG	II	✓	L2	Patients imagine pictures viewed earlier in the same session; no exemplar during the imagery period
Harrison & Tong 2009 [27]	fMRI	II	✗	L1	Delay-period decoding of a just-seen grating (boundary case by design)

Three verdicts of Tables 3 and 4 are justified in this note rather than in rows. The two EEG datasets of [85, 86], examined trial by trial from their data descriptors: [85] is ✗ · L1, since each concept is perceived and then imagined within the trial, the auditory condition asking for imagery in the voice of the speaker presented; [86] is ✗ · L3, since each trial shows a 0.6 s picture and a mask before the imagery period but participants are told to visualise their own representation of the concept rather than the picture shown (§ 2.2). The preprint [41]† is ✓ · L3, since imagery is generated from short text prompts and no image is shown.

not $\Lambda_{\mathrm{within}}$, hence (✓). [30, 39, 31] report permutation tests against chance but do not make their relation to a test-time mismatch through the deployed pipeline checkable, hence (✓). S2. [31] evaluates categories absent from decoder training by construction; [29, 37] evaluate artificial shapes with decoders trained only on natural images; [30] identifies against large sets of previously unseen photographs. The NSD-Imagery entries are (✓), since their test stimuli are held out from training but five of six complex stimuli come from NSD shared1000 and the corpus is cluster-structured (§ 6.3).

S3. Studies that tabulate both a low-level and a semantic metric [29, 28, 17] are scored (✓): S3 also requires the number of alternatives, the distractor-sampling rule and the proportion of same-category distractors (§ 7.4), and these cells are scored from the published descriptions of the distractor protocol rather than a complete re-check against that requirement. (✓) is also given where a single identification metric accompanies a qualitative assessment, and — where the output is a label. S4. [29] collects a five-point vividness rating on every imagery trial and reports its distribution but does not relate it to decoding score, and [31] likewise collects a vividness judgement, so both are (✓). [28] collects a binary vividness response per trial, but because no analysis against it is reported we score ✗. S5, S6. Neither is reported in the main paper or supplement of any of the ten; both require only the generator, the stimulus corpus and predictions already computed, so they could be supplied retrospectively. S7. ✓ where the number of samples per trial and the selection rule are stated and best/median/worst cases are shown [28, 17], or where the reconstruction is deterministic given the data [30]; (✓) where multiple runs are described but the figure-selection rule is not; ✗ where neither is stated. S8. ✗ for [42], where fivefold Gaussian-noise augmentation precedes a 60/20/20 split. Elsewhere ✓: the fMRI studies split by run or session, [5] cross-validates at the trial level and [1] on pseudo-trials averaged from disjoint sets of trials, both within a randomised design. S9. ✓ wherever the trial timeline is reported in enough detail to derive the verdict in Table A.8; ✗ for [42], which reports the timeline but does not address what the in-trial cue video implies for maintenance. S10. ✓ where eye state is specified and monitored [39, 31]; (✓) for [1], where eye state is specified but the alpha-band implication is not analysed; ✗ for [42], which reports eyes open during imagery without a corresponding statement for the cue period or an artefact analysis; — for fMRI studies.

Electrophysiological Imagery Results

Table D.9 gathers the electrophysiological imagery studies of § 5.2 and 6.4, with the class count, chance level and validation unit that § 7.2 requires before an accuracy can be interpreted. Qualitative entries restate findings given in the main text, and admissibility follows Table A.8. Cells are taken from the original reports and data descriptors; where an item could not be located there, the cell says so.

Table D.9: Electrophysiological imagery decoding at a glance. “Validation unit” is the level at which data were split or permuted (standard S8), and “Adm.” the admissibility verdict of § 2.2. No row reports image-level reconstruction scored against an image-level target.

Study	Mod.	Imagined content (classes $k$)	Chance	Reported result	Validation unit	Eye state	Adm.
EEG content decoding
Xie et al. 2020 [1]	EEG, 64 ch	Twelve everyday objects, cued by spoken words after a training phase (12, pairwise)	50%	Above chance in the alpha band; cross-classifies with late-stages of perception	Pseudo-trials averaged from disjoint trials, leave-one-pseudo-trial-out; sign-permutation cluster test	Fixation cross in both tasks; alpha implication not analysed	✓ · L2
Stecher & Kaiser 2024 [2]	EEG, 64 ch	Sixteen scenes from three-sentence descriptions (16, mean pairwise)	50%	Individual scenes and graded scene properties decoded from alpha power	Leave-one-trial-per-scene-out; TFCE with 10,000 sign permutations	Open, fixating a central dot	✓ · L3
Shatek et al. 2019 [5]	EEG, 64 ch	One of four in-trial images (two categories, two exemplars each), retro-cued by serial position (2)	50%	No above-chance decoding during imagery at any time point; perception decoded robustly	Leave-two-trials-out, one per category; per-participant label permutation (1,000)	Open; fixation cross, imagery projected into an on-screen frame	✗ · L1
Kwon et al. 2025 [42]	EEG, 64 ch	Upper-limb actions after an in-trial cue video (4)	25%	72.3% mean accuracy; leave-one-subject-out also reported	Augmented trial (augmentation before a 60/20/20 split)	Eyes open in imagery; cue period not stated	✗ · L1
Shimizu & Srinivasan 2022 [45]	EEG	Images shown for 500 ms before imagery (40)	2.5%	GAN images from imagery-period EEG classified slightly above chance by a ResNet	500 ms segment within an episode	Open (instructed)	✗ · L1

Study	Mod.	Imagined content (classes $k$)	Chance	Reported result	Validation unit	Eye state	Adm.
EEG datasets and BCI control
Gao et al. 2026 [32]	EEG, 32 ch	Ten pictures in three categories, each shown for 4 s in-trial; classified within category (3, 3, 4)	33%, 33%, 25%	Dataset, 22 participants in two sessions; VVIQ administered; EEGNet 62.0% (objects) to about 75% (animals, figures), CSP + KNN 20–26 points lower	Random within-subject 80/20 split of trials, per session	Not stated in the sources examined	✗ · L1
Wilson et al. 2023 [85]	EEG, 124 ch	Three concepts (penguin, guitar, flower) in pictorial, orthographic and auditory form, each imagined after being perceived in the trial (3)	50% (task)	Dataset, 12 participants; technical validation decodes imagination versus perception (75% for the visual modalities), not content	Stratified 5-fold cross-validation over trials, 50 repetitions	Not specified; room darkened	✗ · L1
Rybar et al. 2025 [86]	EEG, 64 ch; fNIRS	Animals versus tools (18 concepts each), visualised after a 0.6 s in-trial picture and mask, alongside silent naming, auditory and tactile imagery (2)	50%	Dataset, 12 participants (EEG and fNIRS) plus 7 (EEG, auditory imagery only); technical validation reports evoked responses, not decoding	None reported in the descriptor	Open; eye movements to be minimised	✗ · L3
Kosmyna et al. 2018 [92]	EEG	Flower versus hammer, each imagined after observing the same image in the trial (2)	50%	The two imagined images not separated above chance (about 52%); imagery versus observation 71%, imagery versus rest 77%	10-fold chronological cross-validation, 5-trial margin	Not stated in the sources examined	✗ · L1
Lee et al. 2020 [93]	EEG	Twelve words or phrases plus rest, imagined as visual scenes after an in-trial visual cue (13)	7.7%	40.1 ± 4.2% for visual imagery, with every participant above the confidence level	Not stated in the sources examined	Not stated in the sources examined	✗ · L1

Study	Mod.	Imagined content (classes $k$)	Chance	Reported result	Validation unit	Eye state	Adm.
Kilmarx et al. 2024 [94]	EEG	Four image categories; short-term imagery after the image, long-term imagery after an auditory cue (4)	25%	Observation 42.1%, short-term imagery 30.1%, long-term imagery 26.7%, against a significance threshold of 29.6%	Leave-one-block-out (5-fold)	Not stated in the sources examined	✗ · L1 ✓ · L2
Dijkstra et al. 2018 [43]	MEG, 275 ch	Face versus house (eight exemplars each), retro-cued after both were shown in the trial (2)	50%	Category decodable from 540 ms after the cue; imagery representations later, more sustained and less temporally specific than perceptual ones	5-fold cross-validation over trials; cluster-based permutation (1,000)	Fixation bull’s-eye; eye tracking, eye position not predictive of category during imagery	✗ · L1
Kreiman et al. 2000 [3]	iEEG, single units	Pictures viewed earlier in the session	—	Picture-selective medial-temporal neurons reinstated during imagery	—	Closed during imagery	✓ · L2
Cerf et al. 2010 [4]	iEEG, single units	Fading a two-image hybrid toward one of four pre-identified images	—	Volitional, closed-loop control of pre-identified neurons; target reached on 596 of 864 trials (69.0%)	Online	Open, viewing the hybrid image (stimulus onscreen)	—

— = not applicable; for [4] no nominal chance level applies, since a trial ends when the target is reached or on time-out. Entries without a number restate findings reported in the main text, and no value has been estimated. “Not stated in the sources examined” means the item was not found in the report or data descriptor consulted. The flower–hammer imagery figure for [92] is as summarised in [94]; for [94] the short-term condition shows the image inside the trial and the long-term condition recalls it from memory on an auditory cue.

Inflation of Pairwise Identification under Categorical Structure

This section derives Eq. (21) together with its finite-sample form, which quantify how far a pairwise identification score can exceed chance when the decoder recovers category but no within-category detail [10].

Let $y$ be a test sample, $\hat{y}$ the prediction for it, and $y^{-}$ a candidate drawn from the remaining test samples. Define the indicator

for an arbitrary similarity $\operatorname{sim}(\cdot,\cdot)$. The pairwise identification accuracy over $n$ test samples is

Now model a decoder that recovers semantic category but not appearance. Partition the triplets into the set $Z$ in which test and candidate belong to the same category and its complement $\bar{Z}$, and assume

The first equation says that two samples of the same category are indistinguishable to this decoder, the second that samples of different categories are discriminated with probability $q$.

Suppose the test set comprises $k$ equally sized categories of $n/k$ samples. For a uniformly drawn candidate $y^{-} \ne y$, the probability that it falls in the same category as $y$ is $\pi_{n,k} = (n/k - 1)/(n - 1)$, so taking expectations in Eq. (E.2) and substituting Eq. (E.3) gives the exact finite-sample accuracy

As $n$ grows, $\pi_{n,k} \to1/k$ and

which is Eq. (21) (Fig. E.6).

Three readings follow. Setting $q = 1$, that is perfect between-category discrimination with zero within-category information, gives $\mathrm{Acc} = 1 - 1/(2k)$ in the limit: 75% for two categories, 90% for five and 95% for ten. A reported pairwise accuracy therefore carries little information about within-category fidelity unless $k$ is small and known. Conversely, an observed accuracy on a test set of $n$ samples in $k$ categories is consistent with $q = (\mathrm{Acc} - 0.5\pi_{n,k})/(1 - \pi_{n,k})$ and no within-category information at all. Finally, the small-sample correction is not negligible at the sizes imagery benchmarks use. For NSD-Imagery’s eighteen stimuli in three sets of six ($n = 18$, $k = 3$), $\pi= 5/17 \approx0.29$, so $q = 1$ yields $\mathrm{Acc} \approx85.3\%$ rather than the asymptotic 83.3%; for the twelve simple and complex stimuli scored in Panel B of Table 5 ($n = 12$, $k = 2$), $\pi= 5/11 \approx0.45$ and $\mathrm{Acc} \approx77.3\%$ rather than 75%. These figures apply when distractors are drawn across sets, which should be checked against the distractor rule of each benchmark [28, 17]; when distractors are restricted to a set, the relevant $k$ is the number of semantic categories within it, which neither benchmark defines. The NSD corpus from which the complex stimuli are drawn also resolves into roughly forty semantic clusters [10]. This is why the distractor rule must be reported with S3 (§ 7.4), and why hard within-category distractors are preferable wherever the stimulus set permits them.

Notes Supporting the Main Text

Structural and Empirical Empty Cells of the Admissibility Criterion

The two dimensions of the admissibility criterion (§ 2.2) are not fully independent. Because L1 is defined by presentation inside the trial, the cell ✓ · L1 is empty by definition, not for want of studies. The cells ✗ · L2 and ✗ · L3 are possible in principle, by re-presenting a pre-learned exemplar inside the trial or by showing an exemplar and asking the participant to imagine their own, as the visual-imagery task of the EEG–fNIRS dataset of [86] does; they are merely unoccupied in the corpus. Where an exemplar is shown inside the trial, the memory source is therefore L1 unless the instruction directs imagery away from the exemplar shown, and the verdicts of Table 3 and Table 4 and of Figure 3 follow this rule.

Implementation of the Null-Signal Control

The normalised neural dependence $\widetilde{\Lambda}$ of Eq. (17) is defined when its denominator is positive. It is not confined to $[0,1]$: it exceeds 1 when the null scores below chance, which flags a bias in the operator itself of the kind displayed by the below-chance CLIP scores of Table 5, Panel B, and it is negative when real input scores below the null; it is therefore reported unclipped. Test-time mismatch may also use phase-scrambled activity in place of another trial’s activity.

Label outputs. For label outputs ($G = \operatorname{id}$) a mismatched trial is scored against the label of the trial it replaces, so its expected score is $\mathcal{E}_{\mathrm{chance}}$ by construction; $\Lambda$ then equals the above-chance margin, $\widetilde{\Lambda} = 1$, and the control reduces to a valid permutation test of held-out accuracy.

Permutation unit and cost. Permutations are drawn at the level of the imagery episode (S8), and the cost of the null re-runs is bounded by re-running only the stages downstream of the substituted input with cached seeds.

Cue leakage. One leak survives every mismatch null: where a stimulus-specific cue remains on screen during imagery, as the single-letter cues of NSD-Imagery do, an imagery-trained decoder can obtain $\Lambda_{\mathrm{within}} > 0$ from the response to the cue. Such decoders should add a control decoded from a cue-only window or under a reassigned cue–stimulus mapping.

The Two Design Axes

Fig. F.7 places the families by their direction of inference and kind of generator (§ 4.1). The caption of Figure 1, which describes Family I as inverting $g_{\phi}$, applies strictly to Ia, and Table 3 keeps the historical label “encoding-model inversion through a feature space” for Ib; both are to be read under these axes.

The Causal Chain of Spurious Reconstruction

Fig. F.8 summarises the chain described in § 7.3: a clustered corpus induces output dimension collapse, collapsed latents are rendered as fluent class-typical images, categorical test structure and fluency-rewarding raters inflate their scores, and the score is read as fidelity. The same chain applies at deployment, where a hallucinated reconstruction could mislead a closed-loop user (§ 8.4) or be misread as evidence (§ 9.4).

A Within-Subject Test of the Nature of the Cross-State Shift

Fig. F.9 outlines the analysis proposed in § 8.2. Reporting transfer error in both directions and the alignment of the weight maps, in addition to cross-validated prediction of the proxy latent, would make the comparison between the gain-only and refitted models interpretable region by region.

The Imagery Corpus by Year and Modality

Fig. F.10 arranges the corpus of Table A.8 and the closed-loop study of Table D.9 by year and modality. Every image-scored reconstruction lies in the fMRI column (§ 6.4).