Mirage of Synthesis: Cross-Domain Surrogates

• Mirage of Synthesis is a cross-domain motif defining surrogate representations that replace direct synthesis with latent, spectral, or semantic substitutes.
• It appears in optics through spectral encoding, in biomedical imaging via latent distillation, and in crystal generation as flexible atom placeholders.
• The concept enables efficient generation in video systems and robust forensics, while also informing theories of semantic equilibrium in agentic machine intelligence.

“Mirage of Synthesis” appears in recent arXiv literature as a recurring label for technically distinct mechanisms of representation construction, reconstruction, and semantic detachment. In optics, it denotes a virtual image recovered from a spectrally encoded one-dimensional source; in biomedical machine learning, it denotes cross-modal latent distillation that substitutes for direct MRI synthesis; in crystal generation, it denotes atom slots that can alternate between existent and non-existent states; in video world models, it denotes latent spatial memory that replaces explicit RGB-space caches; in synthetic-image forensics, it denotes a benchmark centered on visible but underdetected artifacts; and in agentic machine intelligence, it denotes an equilibrium concept for causally detached semantic states (Grusche, 2014, Wu et al., 2 Mar 2026, Okhotin et al., 18 Nov 2025, Wang et al., 8 Jun 2026, Sharma et al., 4 Oct 2025, Tembine, 2 Jun 2026, Sundararaman et al., 9 Jun 2025). A plausible implication is that the phrase functions less as a single doctrine than as a cross-domain motif for synthesis mediated by latent, spectral, or semantically transformed surrogates rather than direct, fully grounded reconstruction.

1. Terminological range and recurrent structure

Across the cited works, “Mirage” names systems that replace direct formation of a target object with an intermediate representation that is easier to manipulate, propagate, or regularize. In some cases the surrogate is explicitly physical, as in spectral encoding on a one-dimensional filament. In others it is latent, as in cross-cohort MRI distillation or latent-space video memory. In still others it is semantic and potentially detached from external reality, as in Causal Mirage Equilibrium.

Domain	Paper	Mirage construct
Optics	“Spectral synthesis provides 2-D videos on a 1-D screen with 360°-visibility and mirror-immunity” (Grusche, 2014)	Virtual 2-D image recovered from a 1-D spectrally encoded screen
AD diagnosis	“MIRAGE: Knowledge Graph-Guided Cross-Cohort MRI Synthesis for Alzheimer’s Disease Prediction” (Wu et al., 2 Mar 2026)	Anatomy-guided cross-modal latent distillation
Crystal generation	“MiAD: Mirage Atom Diffusion for De Novo Crystal Generation” (Okhotin et al., 18 Nov 2025)	Type-0 mirage atoms that permit changing atom count
Video world models	“Latent Spatial Memory for Video World Models” (Wang et al., 8 Jun 2026)	Latent cache of 3D scene tokens
Synthetic-image forensics	“Mirage: Unveiling Hidden Artifacts in Synthetic Images with Large Vision-LLMs” (Sharma et al., 4 Oct 2025)	Dataset of AI-generated images with visible artifacts
Agentic machine intelligence	“Causal Mirage Equilibrium in Agentic Machine Intelligence” (Tembine, 2 Jun 2026)	Stable manifold of causally detached semantic representations
Audio-to-video generation	“Seeing Voices: Generating A-Roll Video from Audio with Mirage” (Sundararaman et al., 9 Jun 2025)	Audio-conditioned video generation in latent space

This shared naming pattern does not imply methodological identity. The optical system is governed by geometric optics; the biomedical and video systems are latent-variable generative or reconstructive pipelines; the crystal system is a diffusion model over variable-cardinality structures; the forensic dataset is evaluative; and the equilibrium concept is formal game theory. What recurs is a structural substitution: a line for a screen, a latent for a volume, a type-0 slot for a fixed atom count, a latent cache for RGB point clouds, artifact explanations for opaque detection, or semantic self-reinforcement for causal grounding.

2. Optical origin: spectral synthesis and projected-image circumlineascopy

The most literal “mirage” in this set is the optical construction of Projected-Image Circumlineascopy (PICS), where a standard LCD/video projector, a first dispersive element, a one-dimensional translucent screen, and a second dispersive element produce semitransparent, rainbow-coloured, virtual 2-D videos that face every viewer anywhere around the 1-D screen and remain invariant under reflection of the 1-D screen in mirrors parallel to it (Grusche, 2014). The projector displays either two narrow white vertical lines for calibration or a 2-D grayscale frame bounded between them. A prism or transmission grating disperses the projection so that the two Newtonian spectra “kiss” at a narrow locus $L_s$, where a thin translucent filament is placed.

The synthesis geometry is expressed by the dispersive displacement

$s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$

with the kissing condition

$|AB-CD| = |s|.$

At $L_s$, each pixel column of the grayscale pattern modulates the intensity of a narrow band around $\lambda_s(x)$. Viewing that one-dimensional rainbow stripe through the second dispersive element produces a virtual image with width and height

$W_a = |a| + w, \qquad h_a = h_s,$

where

$a = 2\,\tan\!\bigl(\tfrac12\,\delta_a\bigr)\,d_a.$

The paper also derives the spectral bandwidth per pixel and the number of resolvable pixels across the reconstructed image. Its reported practical limits are equally explicit: light efficiency is low, typically $<1\%$ of projector light ends up in the single stripe, and there is a trade-off between spectral resolution and spatial resolution (Grusche, 2014). A plausible historical reading is that this work provides the most literal template for later “mirage” terminology: a complete-seeming image is not projected in ordinary two-dimensional form but recovered from a constrained encoded surrogate.

3. MIRAGE in Alzheimer’s disease prediction: cross-cohort latent distillation

In “MIRAGE: Knowledge Graph-Guided Cross-Cohort MRI Synthesis for Alzheimer’s Disease Prediction,” the missing-MRI problem is reformulated as an anatomy-guided cross-modal latent distillation task rather than direct voxel synthesis (Wu et al., 2 Mar 2026). The motivating setting is multimodal AD diagnosis combining structural MRI and EHR, where MRI is expensive and frequently unavailable, while EHR is routinely collected but lacks direct anatomical priors. The paper explicitly rejects de novo 3D anatomical reconstruction from sparse tabular records as ill-posed and clinically risky.

The architecture has three stages: $(A)$ 3D MRI autoencoding, $(B)$ KG-guided embedding propagation, and $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$0 adapter plus decoder alignment with skip-feature compensation. The Biomedical Knowledge Graph has nodes $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$1, where $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$2 are patients with real MRI, $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$3 are patients without MRI, and $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$4 are medical-concept nodes. Patient–concept edges are created by embedding each EHR concept description with SapBERT (768-D) and linking patient nodes to top-$s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$5 semantically matching concept nodes, while concept–concept edges and preexisting medical relations are inherited from iBKH with $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$6M entities and $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$7M triples. Initial node features are $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$8 for MRI-available patients, SapBERT concept embeddings $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$9 for medical concepts, and $|AB-CD| = |s|.$0 for MRI-missing patients.

Message passing is performed by a three-layer GATv2 with residuals, trained first by supervised regression onto real-MRI latents,

$|AB-CD| = |s|.$1

and then by decoder-consistency fine-tuning using a frozen pre-trained 3D U-Net decoder,

$|AB-CD| = |s|.$2

The decoder is used strictly as an auxiliary regularization engine. MRI-missing patients lack their own high-frequency skip features, so the method retrieves them from top-$|AB-CD| = |s|.$3 clinically similar $|AB-CD| = |s|.$4 neighbors in latent space; in practice $|AB-CD| = |s|.$5 suffices. The resulting cohort-aggregated skip feature compensation provides what the paper calls a plausible anatomical canvas for the frozen decoder.

The central operational claim is that MIRAGE completely bypasses computationally expensive 3D voxel reconstruction at inference. For a new EHR-only patient, the system links the patient into the KG, runs the GAT to obtain $|AB-CD| = |s|.$6, applies whitening-coloring alignment and an adapter $|AB-CD| = |s|.$7 to obtain $|AB-CD| = |s|.$8, discards the decoder and skip features, and feeds $|AB-CD| = |s|.$9 directly into the classification head. On ADNI, with $L_s$0 subjects, $L_s$1-dim EHR, $L_s$2 MRI volumes, CN vs. AD labels, and a $L_s$3 train/test split with test MRIs masked, MIRAGE + synthetic-MRI fusion yields a $L_s$4 absolute gain in balanced accuracy over the strongest EHR-only baseline; ablations show that KG, GAT, AE prior, and adapter are each essential, and human expert evaluation confirms anatomical realism and clinical usability of the structural surrogate (Wu et al., 2 Mar 2026).

4. Mirage infusion in crystal diffusion

In “MiAD: Mirage Atom Diffusion for De Novo Crystal Generation,” the mirage construct is a special atom type rather than an optical or latent surrogate (Okhotin et al., 18 Nov 2025). Traditional crystal diffusion models fix the number of atoms $L_s$5 in the unit cell at sampling time. Mirage infusion lifts this restriction by introducing a special atom-type, type $L_s$6, called a mirage atom. Each crystal $L_s$7 of $L_s$8 real atoms is expanded into an augmented crystal

$L_s$9

containing $\lambda_s(x)$0 positions, of which $\lambda_s(x)$1 are real and the remainder are mirages. Mirage atoms begin as placeholders at random positions and during backward diffusion may materialize into real types or vanish back into type $\lambda_s(x)$2.

The joint diffusion model factorizes over lattice, fractional coordinates, and atom types, with a D3PM over $\lambda_s(x)$3 types including the mirage state. The mirage variable is read directly from the type assignment,

$\lambda_s(x)$4

Training uses a single neural network with total loss

$\lambda_s(x)$5

where the coordinate score-matching term is masked so that only real atoms contribute, while the discrete KL over atom types includes both real and mirage atoms.

The sampling procedure explicitly avoids fixing the number of real atoms at $\lambda_s(x)$6. Each slot is initialized from $\lambda_s(x)$7, the backward process may flip $\lambda_s(x)$8 or $\lambda_s(x)$9, and only after $W_a = |a| + w, \qquad h_a = h_s,$0 are type-$W_a = |a| + w, \qquad h_a = h_s,$1 slots dropped by Reduction. Quantitatively, on MP-20 with $W_a = |a| + w, \qquad h_a = h_s,$2 samples and DFT-computed stability, MiAD achieves $W_a = |a| + w, \qquad h_a = h_s,$3 stability $W_a = |a| + w, \qquad h_a = h_s,$4, $W_a = |a| + w, \qquad h_a = h_s,$5 Unique & Novel, and $W_a = |a| + w, \qquad h_a = h_s,$6 S.U.N., versus $W_a = |a| + w, \qquad h_a = h_s,$7, $W_a = |a| + w, \qquad h_a = h_s,$8, and $W_a = |a| + w, \qquad h_a = h_s,$9 for DiffCSP. M.S.U.N. improves from $a = 2\,\tan\!\bigl(\tfrac12\,\delta_a\bigr)\,d_a.$0 to $a = 2\,\tan\!\bigl(\tfrac12\,\delta_a\bigr)\,d_a.$1, and similar $a = 2\,\tan\!\bigl(\tfrac12\,\delta_a\bigr)\,d_a.$2–$a = 2\,\tan\!\bigl(\tfrac12\,\delta_a\bigr)\,d_a.$3 gains hold under MLIP-based stability estimates. The key ablation is direct: turning mirage infusion off yields S.U.N. $a = 2\,\tan\!\bigl(\tfrac12\,\delta_a\bigr)\,d_a.$4, and adding mirage infusion with identical architecture yields S.U.N. $a = 2\,\tan\!\bigl(\tfrac12\,\delta_a\bigr)\,d_a.$5 (Okhotin et al., 18 Nov 2025).

Here the “mirage” is not an error mode but a controlled state variable that allows nonmonotonic walks in composition and structure. A plausible implication is that the paper extends the mirage metaphor from perceptual appearance to cardinality-flexible generative search.

5. Latent-space synthesis in video systems

Two 2025–2026 papers use “Mirage” for video generation, but they do so in different ways. “Latent Spatial Memory for Video World Models” introduces a persistent 3D cache that stores scene information directly in the diffusion latent space and names the resulting framework Mirage (Wang et al., 8 Jun 2026). Instead of an explicit RGB point-cloud memory $a = 2\,\tan\!\bigl(\tfrac12\,\delta_a\bigr)\,d_a.$6, the method stores

$a = 2\,\tan\!\bigl(\tfrac12\,\delta_a\bigr)\,d_a.$7

Latent grid cells are lifted into 3D by depth-guided back-projection,

$a = 2\,\tan\!\bigl(\tfrac12\,\delta_a\bigr)\,d_a.$8

and read back out for a target camera by projection and z-buffering, assigning each latent cell the frontmost point feature $a = 2\,\tan\!\bigl(\tfrac12\,\delta_a\bigr)\,d_a.$9. This removes repeated full-resolution rasterization and VAE re-encoding from the conditioning loop. The reported gains are up to $<1\%$0 faster end-to-end video generation and $<1\%$1 reduction in memory footprint relative to explicit 3D baselines. On WorldScore, Mirage achieves average score $<1\%$2, with Static/Dynamic $<1\%$3, 3D Consistency $<1\%$4, and Photo Consistency $<1\%$5; on RealEstate10K it reports PSNR $<1\%$6, SSIM $<1\%$7, LPIPS $<1\%$8, closed-loop PSNR$<1\%$9 $(A)$0, SSIM$(A)$1 $(A)$2, and LPIPS$(A)$3 $(A)$4 (Wang et al., 8 Jun 2026).

“Seeing Voices: Generating A-Roll Video from Audio with Mirage” instead uses Mirage as an audio-to-video foundation model for realistic, expressive A-roll generation from scratch given audio input (Sundararaman et al., 9 Jun 2025). The model builds on a $(A)$5B-parameter Diffusion Transformer backbone, the “Mochi” variant, with a 3D spatiotemporal VAE, frozen wav2vec audio encoder, optional T5-XXL text encoder, optional reference-image pathway, and $(A)$6 asymmetric self-attention Transformer blocks. A $(A)$7 s, $(A)$8-frame, $(A)$9p clip is downsampled to latent size $(B)$0, yielding a flattened video-token sequence length of approximately $(B)$1. Training uses a flow-matching objective,

$(B)$2

with classifier-free guidance enabled by random modality dropping. The preprocessing pipeline includes scene detection with PySceneDetect, single-speaker cropping via MediaPipe, OCR filtering with CRAFT + EasyOCR, logo filtering with a YOLOX-based detector, and lip-sync filtering with SyncNet. The paper states that no GAN and no explicit cross-modal alignment loss are used; qualitative findings include plosive and viseme precision, natural blinking and gaze shifts, emotion-driven expressions, paralinguistic events, and gesture–speech semantic alignment. It also states that exact numeric benchmark comparisons such as FVD values are not provided, and emphasizes human-centric and qualitative evaluation (Sundararaman et al., 9 Jun 2025).

Taken together, these papers suggest a specific modern usage of “mirage” in video research: high-fidelity synthesis is achieved by operating in the model’s native latent geometry rather than by repeated detours through full pixel-space reconstruction.

6. Forensic and theoretical uses: hidden artifacts and semantic detachment

In image forensics, “Mirage” designates a dataset rather than a generator. “Mirage: Unveiling Hidden Artifacts in Synthetic Images with Large Vision-LLMs” constructs a benchmark of $(B)$3 images, equally split between $(B)$4 real photographs from MS-COCO and $(B)$5 synthetic images from JourneyDB and DALL·E-3, all resized to between $(B)$6 and $(B)$7 pixels (Sharma et al., 4 Oct 2025). Synthetic samples are selected through an LVLM-guided pipeline: Qwen-VL predicts artifact categories $(B)$8, CLIP similarity

$(B)$9

is computed, images with fewer than five predicted artifacts or $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$00 are discarded, and the top $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$01 by $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$02 form the synthetic half. Manual inspection of $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$03 randomly sampled synthetic images confirms $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$04 accuracy in artifact presence. The taxonomy has nine classes, including malformed anatomy, physical inconsistencies, texture anomalies, frequency artifacts, color bleed, semantic incoherence, geometry distortion, edge halos, and background blending. On Mirage, zero-shot Qwen 2.5 B achieves $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$05 accuracy and $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$06 $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$07, outperforming CODE at $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$08 and SPAI at $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$09, while on Chameleon, where no visible artifacts exist, CODE remains highest at $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$10 and VLM$s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$11 drops to $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$12 (Sharma et al., 4 Oct 2025). The paper’s central claim is therefore diagnostic: visible artifacts remain detectable by human-like perceptual models even when specialist detectors tuned to older fingerprints fail.

At the theoretical extreme, “Causal Mirage Equilibrium in Agentic Machine Intelligence” uses the term to formalize endogenous epistemic decoupling (Tembine, 2 Jun 2026). The paper defines a risk-sensitive mean-field-type equilibrium in which semantic representations stabilize into detached but operationally robust manifolds. The key scalar is the dimensionless Mirage Intensity,

$s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$13

where $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$14 is endogenous self-reinforcement, $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$15 is operational confidence, and $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$16 is causal alignment. The interpretation is explicit: $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$17 is the realism regime, $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$18 is the critical bifurcation surface, and $s = 2\,\tan\!\bigl(\tfrac12\,\delta_s\bigr)\,d_s,$19 is the mirage regime. Under continuity, compactness, convexity, concavity, and hemicontinuity assumptions, the paper proves existence of a Causal Mirage Equilibrium by the Kakutani-Glicksberg-Fan fixed-point theorem on the space of joint probability measures. It also states a non-linear mirage bifurcation theorem: when endogenous reinforcement dominates causal grounding, the grounded fixed point becomes unstable and a stable invariant manifold of ungrounded states bifurcates. The examples given are hallucinated citations in large-LLMs, multi-agent generative pipelines with memory loops, and risk-sensitive planning agents that ignore rare contradictory evidence (Tembine, 2 Jun 2026).

A plausible synthesis of the forensic and theoretical strands is that they bracket two levels of the same broad concern. The forensic work studies where synthesis still leaves observable traces; the equilibrium work studies when internally coherent synthetic semantics no longer require such grounding at all.