Hyperphantasia: Vivid Imagery Spectrum

• Hyperphantasia is a multisensory vivid imagery phenotype marked by exceptionally strong top‐down reactivation across various sensory modalities.
• Research using scalar vividness ratings and Ganzflicker-induced paradigms shows that strong imagers generate qualitatively richer, semantically structured internal content.
• Neuroimaging and machine-learning benchmarks suggest that hyperphantasia involves efficient sensory-cognitive integration and differentiated neural network connectivity.

Searching arXiv for papers on hyperphantasia and related imagery research. arXiv search query: hyperphantasia imagery agnosia Ganzflicker benchmark multimodal LLMs Hyperphantasia denotes extremely vivid mental imagery and is commonly positioned at the vivid end of a continuum whose opposite pole is aphantasia. In a neuropsychological framing proposed by Włodzisław Duch, however, hyperphantasia is not best understood as a narrowly visual anomaly but as one endpoint of a broader spectrum of imagery agnosia, defined in terms of the strength or weakness of top-down reactivation of sensory qualia across modalities (Duch, 2021). More recent work has refined this picture by showing that strong imagers differ not only in rated vividness but also in the semantic and sensorimotor structure of internally generated content under induced-hallucination paradigms, while a separate line of machine-learning research has appropriated the term “Hyperphantasia” for a benchmark of mental visualization in multimodal LLMs rather than for the human phenotype itself (Chkhaidze et al., 11 Jul 2025, Sepehri et al., 16 Jul 2025).

1. Conceptual definition and scope

Duch places hyperphantasia at the extreme vivid end of individual differences in mental imagery, with aphantasia at the opposite end, but argues that the prevailing focus on visual imagery is too narrow. In his account, imagery differences are “much more widespread,” extending to auditory, tactile, olfactory, gustatory, kinesthetic, and bodily imagery. He therefore proposes the broader neuropsychological umbrella term “imagery sensory agnosia”, under which visual aphantasia and hyperphantasia are treated as modality-specific instances of a more general continuum of sensory imagery ability (Duch, 2021).

This framing shifts the object of analysis from introspective vividness alone to the capacity for top-down re-creation of conscious qualia. Hyperphantasia, on this view, is not merely “strong imagination” in an informal sense. It denotes unusually effective access to sensory-like internal reactivation. Duch explicitly states that this should hold “in relation to all sensory modalities,” and notes that vividness distributions differ across modalities and are only partially correlated. Some respondents with low visual imagery report strong imagery in another modality, which argues against reducing hyperphantasia to a unitary visual trait (Duch, 2021).

A central misconception is therefore that hyperphantasia is simply the visual opposite of aphantasia. Duch’s framework suggests a more differentiated interpretation: vivid imagery can be modality-specific, multisensory, or dissociated from other aspects of cognition such as memory, recognition, and affect. A plausible implication is that hyperphantasia is better treated as a family of imagery phenotypes than as a single homogeneous condition.

2. Spectrum models and operationalization

The visual imagery spectrum is operationalized explicitly in the 2025 Ganzflicker study through a 0–10 self-rating of visual imagery vividness, where 0 corresponds to no mental imagery and 10 to imagery “as vivid as real perception.” For group analyses, scores are binned into weak imagers (0–3), moderate imagers (4–7), and strong imagers (8–10), with the strong group intended to approximate hyperphantasia / vivid imagery. In the analyzed dataset, group sizes were Weak = 1,515, Moderate = 1,634, and Strong = 1,216 (Chkhaidze et al., 11 Jul 2025).

This operationalization is deliberately scalar, but the paper’s central claim is that the upper end of the spectrum should not be reduced to a scalar increase in vividness. Strong imagers are argued to show qualitatively richer, more semantically structured internal content during Ganzflicker-induced hallucinations. The study therefore treats vividness ratings as a starting point rather than a complete characterization of hyperphantasia.

The same paper formalizes several analyses around this spectrum. Topic-based prediction of vividness is expressed as

$$\text{vividness}_i = \beta_0 + \sum_{k=1}^{27} \beta_k \, \text{topic}_{ik} + \epsilon_i$$

with an L1 penalty, and representational dissimilarity across vividness bins is defined as

$D_{ij} = |v_i - v_j|.$

These formulations matter because they make explicit that hyperphantasia can be modeled both as a continuous trait and as an endpoint phenotype. The paper’s failure to classify the moderate group well, despite better-than-chance classification of weak and strong groups, further suggests that the center of the distribution is heterogeneous whereas the extremes are more structurally distinctive (Chkhaidze et al., 11 Jul 2025).

3. Multimodal phenomenology, memory, and dissociation

Although hyperphantasia occupies the vivid end of the continuum, Duch’s phenomenological analysis is centered on the opposite pole, especially auditory imagery agnosia, in order to clarify what imagery consists in. His first-person report after more than 20 years of learning music describes BAIS ratings “consistently at the ‘no image’ bottom of the scale.” Questionnaire prompts do not evoke inner sound but only a “feeling that I know these sounds and will recognize them when I hear them.” The distinction between a sensory image and a mere feeling of knowing is one of the paper’s key conceptual discriminations (Duch, 2021).

This dissociation is theoretically relevant to hyperphantasia because it shows that imagery is separable from recognition, categorization, motor skill, and procedural memory. Duch reports that he can recognize instrument timbre automatically, learn fingering patterns, and perform from score, yet cannot anticipate how an instrument will sound before playing it. He can sometimes recall sequences of finger movements without access to sensory imagery; he writes that melodies may be stored in long-term and procedural memory “although not accessible to sensory imagery.” Hyperphantasia, by implication, is not identical with superior memory or superior skill. It concerns accessibility of sensory qualia, not the totality of musical, mnemonic, or perceptual competence.

The paper also argues for the possibility of covert imagery without conscious sensory qualia. Duch suggests that his improvisation is aided by such covert imagery even though he lacks inner audition, asking whether it makes sense to speak of “hidden or unconscious imagery.” This aligns imagery agnosia with analogies such as blindsight and color anomia, while also emphasizing that these are only analogies: bottom-up recognition and top-down imagery are not symmetric systems. A plausible implication is that hyperphantasia should not be defined solely by introspective confidence; the relevant contrast is between stronger or weaker access to consciously recreated sensory states.

Duch further distinguishes perception, memory, and imagery as “three partially independent dimensions,” and adds emotion as a fourth. He treats autobiographical recall as related but not identical to imagery vividness. His own autobiographical memory is described as verbally mediated and event-based rather than vivid “mental time travel,” often cued by photographs. This matters for hyperphantasia because vivid imagery and vivid autobiographical re-experiencing may covary without being identical constructs (Duch, 2021).

4. Putative neural mechanisms

Duch’s mechanistic proposal is that imagery requires top-down reactivation of sensory cortices from associative and memory systems. He summarizes this directly: visual, auditory, and other sensory imagery needs to recreate conscious qualia by reactivating sensory cortices in a top-down manner. Within this framework, hyperphantasia corresponds to unusually strong top-down activation and/or especially responsive sensory cortex, whereas imagery agnosia reflects weak or ineffective top-down activation (Duch, 2021).

The paper links this account to differences in network architecture and excitability. Strong imagery is noted to be associated with stronger prefrontal-visual network connectivity in resting-state fMRI, and Duch cites evidence that “the strength of sensory imagery is predicted by lower neural resting activity and excitability levels.” He further proposes that individual differences depend on the density and strength of top-down and bottom-up projections between primary sensory cortices and associative cortex, together with projections to limbic structures involved in valence. His schematic links Sensory cortices, Memory, Imagery, Appreciation, and Displeasure by information flow, with sensory transduction and limbic valuation contributing to music and art experience (Duch, 2021).

The 2025 Ganzflicker paper offers a compatible but more explicitly layered visual account. It interprets simple geometric hallucination content as potentially supported by early visual cortex, while more complex naturalistic content may require higher-order regions such as fusiform and scene-related areas, coordinated by frontoparietal control networks. On this interpretation, weak imagers may retain low-level feature representations but show reduced top-down integration into coherent images, whereas strong imagers may exhibit more effective coordination across visual and control systems, enabling richer and more semantically integrated content (Chkhaidze et al., 11 Jul 2025).

Taken together, these accounts suggest that hyperphantasia is not reducible to stronger activation in a single cortical locus. A plausible implication is that it reflects more efficient coupling between sensory reinstatement, associative structure, and control processes that stabilize complex internally generated scenes.

5. Experimental characterization beyond vividness

The strongest recent empirical refinement of hyperphantasia comes from the Ganzflicker study. Ganzflicker is a rapidly alternating red-and-black display, presented at 7.5 Hz for approximately 10 minutes under dim lighting with white noise, and is reported to induce hallucinations in over 80% of participants in large-scale studies. The authors analyzed an open dataset in which 6,664 individuals initially underwent the task; after exclusions, the final corpus comprised 4,365 participants, with 4,057 retained for the sensorimotor analysis and over 10,000 sentence-level documents used for topic modeling (Chkhaidze et al., 11 Jul 2025).

Using BERTopic with Sentence-BERT embeddings, UMAP (n_components = 10, n_neighbors = 15, min_dist = 0.1), HDBSCAN (min_cluster_size = 30), c-TF-IDF / BM25+, and GPT-4o-mini for topic labeling, the study identified 27 non-outlier topics with topic coherence C = 0.50. The extracted themes ranged from simple geometric or perceptual content—lines, spirals, color flashes, fractals, visual patterns—to dynamic spatial imagery and naturalistic scenes such as faces, forests, city skylines, space travel, ocean-at-night, and butterflies. Strong imagers reported more complex, naturalistic, semantically meaningful content, whereas weak imagers reported more simple geometric forms and were characterized not only by simplicity but by the absence of structured/naturalistic content (Chkhaidze et al., 11 Jul 2025).

The predictive analyses reinforce this distinction. A Lasso regression using the 27 topic probabilities yielded R² = 0.05, MSE = 8.63, α = 0.03, with 21 of 27 topics retained with non-zero coefficients. Positive predictors of vividness included morphing faces (B = 0.29), moving hallways (B = 0.14), stars & galaxies (B = 0.14), city skylines (B = 0.13), and forest scenery (B = 0.13), while negative predictors included visual patterns (B = -0.21), color flashes (B = -0.11), and lines (B = -0.11). In one-vs-rest Lasso-regularized logistic regression, the weak imagery classifier reached F1 = 0.54, p < .001, the strong imagery classifier F1 = 0.44, p = .002, and the moderate imagery classifier F1 = 0.43, p = .57. The strong group retained 21 of 27 topics, whereas the weak group retained 10 of 27, supporting the view that hyperphantasia involves a broader and more differentiated content repertoire rather than simply more intense imagery (Chkhaidze et al., 11 Jul 2025).

The embedding analyses show that this graded structure is more closely captured by multimodal representations than by text-only semantics. Comparing representational dissimilarity matrices to the theoretical vividness RDM, the paper reports for text-only models BERT: $r = .67, p < .001$, GPT-2: $r = .40, p = .002$, and RoBERTa: $r = .30, p = .03$, and for vision-LLMs CLIP: $r = .76, p < .001$, SigLIP: $r = .71, p < .001$, and BLIP: $r = .03, p = .83$. The authors interpret this as evidence that imagery-related language carries perceptual structure that some vision-LLMs detect better than text-only models (Chkhaidze et al., 11 Jul 2025).

A further refinement comes from the Lancaster Sensorimotor Norms. Across 13 dimensions—6 perceptual, 5 motor, and 2 composite measures—stronger imagery was associated with richer perceptual and motor grounding. In the composite GLM, perceptual strength predicted vividness with B = 0.35, SE = 0.05, t = 6.54, p < .001; motor strength with B = 0.21, SE = 0.05, t = 4.17, p < .001; and description length with B = 0.42, SE = 0.05, t = 8.39, p < .001. Modality-specific effects remained robust for visual, haptic, olfactory, and auditory dimensions after mediation analysis. The study therefore argues that hyperphantasia is associated not only with vivid visual content but with richer embodied and multisensory references (Chkhaidze et al., 11 Jul 2025).

6. Implications, controversies, and adjacent uses of the term

The literature summarized here supports three corrections to common oversimplifications. First, hyperphantasia is not adequately defined as a purely visual condition; Duch’s account treats it as a multisensory vivid imagery phenotype. Second, it should not be conflated with autobiographical reliving, general memory ability, or perceptual recognition, because the relevant systems can dissociate. Third, recent computational evidence suggests that the strong-imagery end of the spectrum is not merely “more vivid” but structurally different in its internally generated content (Duch, 2021, Chkhaidze et al., 11 Jul 2025).

Duch emphasizes broader implications for consciousness research, metacognition, individual differences, education, and the understanding of mathematical and STEM performance. He notes associations from questionnaire work in which imagery agnosia appears more common in abstract professions, whereas “hyperphantasia or extremely vivid imagery is more frequent in arts, media, and entertainment.” He does not present this as a strict rule, but as an association consistent with the hypothesis that strong imagery may support artistic and musical talent while weak imagery may be more common in abstract, conceptual domains (Duch, 2021).

An adjacent but distinct development is the use of “Hyperphantasia” as the title of a synthetic benchmark for mental visualization in multimodal LLMs. That benchmark contains four synthetic puzzles—Seven Segments, Connect the Dots, Linear Trajectory, and Parabolic Trajectory—organized into Interpolation and Extrapolation, with three difficulty levels and 1,200 total samples. It is designed to test whether models can internally construct missing or future visual structure rather than merely recognize what is directly shown. The benchmark reports a substantial human–model gap and explores improvement via GRPO reinforcement learning on Qwen-VL 2.5 7B, where training on medium-difficulty or mixed-difficulty data generalizes better than training only on easy data (Sepehri et al., 16 Jul 2025).

This machine-learning usage should not be confused with the human phenotype. The benchmark name is evocative rather than diagnostic. Even so, its existence is conceptually noteworthy: it reflects a growing tendency to treat hyperphantasia not only as a subjective vividness trait but as a model case of active internal visualization, involving interpolation, extrapolation, and the construction of latent visual structure. A plausible implication is that future research may increasingly connect human imagery differences with computational accounts of internal simulation, while still requiring careful separation between psychometric, phenomenological, and benchmark-based uses of the term.