Latent-Space Diffusability: Methods & Metrics

• Latent-space diffusability is defined as a latent representation’s capacity to support stable, semantically coherent diffusion dynamics while preserving class structure and spectral properties.
• Methodologies to assess diffusability include geometric measures like OLE, spectral analyses such as ESM/DSM, and trajectory-based controls that ensure robust semantic traversals.
• Applications span from image, text, and video generation to weather forecasting and graph reconstruction, informing tokenizer design and diffusion scheduling for improved generative outcomes.

Latent-space diffusability denotes the extent to which a latent representation supports stable, learnable, and semantically faithful diffusion or flow dynamics. Across recent work, the term is operationalized in several related ways: as preservation of low-dimensional class structure during self-consuming retraining, as robustness of latent edits under iterative denoising, as balanced frequency exposure in high-dimensional latent diffusion, as semantic predictability of initial noise seeds, and as the smoothness and conditioning of continuous latent trajectories for text, graphs, weather fields, and reduced-order dynamical systems (Cai et al., 16 Nov 2025, Zhong et al., 26 Sep 2025, Lai et al., 27 Nov 2025, Zhong et al., 2 Jun 2026). This suggests that latent-space diffusability is not a single invariant scalar, but a family of criteria linking latent geometry, spectral structure, and denoising dynamics to downstream generative performance.

1. Definitions and conceptual scope

In self-consuming image diffusion, latent-space diffusability is defined as a diffusion model’s ability to traverse latent manifolds and produce trajectories that remain close to class-specific low-rank subspaces of real data; high diffusability requires well-separated class subspaces, low OLE, and high per-image confidence under a probing classifier (Cai et al., 16 Nov 2025). In latent editing for Stable Diffusion, it is defined as the propensity of perturbations and traversals applied to internal latent representations to propagate through the denoising process and yield semantically coherent, high-fidelity generations across directions and magnitudes of change (Zhong et al., 26 Sep 2025). In high-dimensional latent diffusion for images, it is defined as the ease with which a generative model can learn and synthesize latent embeddings that decode to realistic, detailed images, with particular emphasis on whether exposure is balanced across the frequencies required by the decoder (Lai et al., 27 Nov 2025).

Other domains sharpen different aspects of the same concept. In diffusion seed-space analysis, latent-space diffusability is the degree to which initial Gaussian seeds admit reliable mappings to semantic attributes, such as class labels, without modifying the generator (Wei et al., 5 Feb 2026). In text diffusion, it refers to how amenable a continuous latent representation of discrete text is to a time-continuous diffusion trajectory that both progressively destroys information in a controlled, data-aligned way and supports a stable reverse-time generative trajectory (Midavaine et al., 7 Jan 2026). In scientific computing, a latent space is “diffusible” when functions of interest and reduced dynamics are sufficiently smooth with respect to diffusion coordinates, so that a second diffusion on latent coordinates can support global interpolation and lifting (Evangelou et al., 2022).

A more tokenizer-centric formulation defines diffusability as how easily latent codes produced by an encoder can be modeled by a downstream diffusion model, emphasizing fast convergence, strong generative quality, and semantic coherence at fixed compute (Zhang et al., 13 May 2026). A more systematic view treats it as the degree to which a latent space reduces the ambiguity and complexity of the velocity field that a flow-matching or diffusion model must learn, which motivates geometric, semantic, spectral, and dynamical predictors such as VIV, LNC, SRSS, and SEC (Zhong et al., 2 Jun 2026).

2. Geometric and subspace formulations

A central geometric formulation studies latent degradation in self-consuming diffusion through a fixed, real-data-trained encoder. For generation $k$, denoising timestep $t$, and sample $x^{(k)}$, the latent representation is

$h^{(k)}_t = e^{(0)}(x^{(k)}, t),$

where $e^{(0)}$ is the encoder of the initial diffusion model trained on real data. For a batch, latent features are assembled into $M^{(k)}_t$, and class-conditional blocks $M^{(k)}_{c,t}$ are used to compute the Orthogonal Low-rank Embedding statistic

$$\mathrm{OLE}^{(k)}_t = \sum_{c \in C} \|M^{(k)}_{c,t}\|_* - \|M^{(k)}_t\|_*.$$

Lower values indicate better separability and more low-rank, orthogonal class subspaces; higher values indicate entanglement and degraded structure. In pure synthetic loops, OLE grows monotonically across generations at fixed $t$, while across timesteps within a generation it follows a U-shaped trajectory, with an intermediate $t$ giving the most separable latent structure (Cai et al., 16 Nov 2025).

The same work formalizes class-conditional latents as noisy low-rank Gaussians,

$t$0

and relates subspace alignment to degradation. If $t$1 denotes a bound on principal angles between class subspaces, the expected OLE satisfies

$t$2

so decreasing subspace angles raises the lower bound and quantifies structural collapse. A complementary result derives a Bayes log-likelihood ratio

$t$3

and shows that expected classifier confidence decreases as subspaces align. The geometric claim is therefore not merely descriptive: lower orthogonality simultaneously raises OLE and lowers confidence, linking latent entanglement to reduced fidelity and diversity under retraining (Cai et al., 16 Nov 2025).

A broader information-geometric treatment defines latent diffusability through the rate of change of MMSE along a Gaussian corruption path. For a $t$4-dimensional space,

$t$5

so

$t$6

where $t$7 is Fisher Information and $t$8 is Fisher Information Rate. In this framework, global near-isometry aligns FI across data and latent spaces, whereas FIR depends on local encoder geometry and decomposes latent distortion into dimensional compression, tangential distortion, and curvature injection. The Hessian transformation

$t$9

makes the distinction explicit: inherited curvature comes from the pullback metric, while encoder nonlinearity injects additional curvature that raises FIR and degrades latent diffusion (Gu et al., 3 Apr 2026).

This geometric line of work implies that diffusability depends both on class- or manifold-level separability and on the local regularity of the encoder-induced metric. A plausible implication is that many empirical tokenizer improvements can be reinterpreted as partial attempts to preserve subspace orthogonality, control curvature injection, or reduce ambiguity in the learned score field.

3. Spectral and frequency-based formulations

A major body of work treats latent-space diffusability as a spectral alignment problem. One influential formulation defines diffusability as how easily a distribution can be modeled by a diffusion process: high diffusability indicates that the distribution is easy to fit, whereas low diffusability makes the process more complex. In image and video latent diffusion, blockwise 2D DCT analysis shows that modern autoencoder latents often have flatter spectra than RGB and exhibit inordinately strong high-frequency components, especially as bottleneck channel size increases. The proposed remedy is scale equivariance, implemented by the decoder-side auxiliary loss

$x^{(k)}$0

with $x^{(k)}$1 and $x^{(k)}$2 obtained by bilinear downsampling. This suppresses spurious high-frequency latent content, aligns latent and RGB spectra, and improves downstream diffusion metrics, including a 19% FID reduction for image generation on ImageNet-1K $x^{(k)}$3 and at least 44% FVD reduction for video generation on Kinetics-700 $x^{(k)}$4 (Skorokhodov et al., 20 Feb 2025).

A complementary frequency analysis of high-dimensional latent spaces argues that generation quality deteriorates at large latent dimensionality because decoders depend strongly on high-frequency latent components, whereas encoders under-represent high-frequency content, particularly when extremely high-frequency RGB signals are present. The resulting under-exposure causes diffusion or flow-matching models to underfit high-frequency latent bands. The proposed FreqWarm curriculum increases early-stage exposure by low-pass filtering RGB inputs before encoding during a warm-up phase and then switching back to full-frequency RGB. Without modifying or retraining the autoencoder, this reduces gFID by $x^{(k)}$5 on Wan2.2-VAE, $x^{(k)}$6 on LTX-VAE, and $x^{(k)}$7 on DC-AE-f32, with larger gains at higher channel counts (Lai et al., 27 Nov 2025).

A more explicitly unifying proposal is the Spectrum Matching Hypothesis, according to which superior latent diffusability requires both Encoding Spectrum Matching (ESM) and Decoding Spectrum Matching (DSM). ESM prescribes a flattened power-law latent PSD,

$x^{(k)}$8

implemented by matching normalized latent PSD to a flattened target image PSD through

$x^{(k)}$9

DSM requires preserving frequency-to-frequency semantic correspondence through shared spectral masks and reconstruction from masked latents. Empirically, DSM-AE achieves the best gFID among the compared variants on CelebA and ImageNet, and the paper argues that prior methods such as VA-VAE, UAE, Scale Equivariance, and EQ-VAE can be interpreted as special cases within the same spectral framework (Ning et al., 15 Mar 2026).

Video latent diffusion extends this spectral perspective beyond spatial frequencies. A statistical analysis of video VAE latents identifies two properties as beneficial for diffusability: a spatio-temporal frequency spectrum biased toward low frequencies and a channel-wise eigenspectrum dominated by a few modes. These are induced by Local Correlation Regularization,

$h^{(k)}_t = e^{(0)}(x^{(k)}, t),$0

and Latent Masked Reconstruction,

$h^{(k)}_t = e^{(0)}(x^{(k)}, t),$1

The resulting Spectral-Structured VAE yields a $h^{(k)}_t = e^{(0)}(x^{(k)}, t),$2 speedup in text-to-video convergence and a 10% gain in video reward, supporting the view that diffusability depends not only on spatial smoothness but also on temporal bias and channel-wise energy concentration (Liu et al., 5 Dec 2025).

Taken together, these works treat the latent spectrum as an interface between tokenizer design and denoiser learnability. The shared conclusion is that latents that are too noisy, too flat, or frequency-misaligned force the diffusion backbone to learn fragile, late-emerging details under poor signal-to-noise conditions.

4. Trajectory ordering, controllability, and semantic traversal

Some formulations of latent-space diffusability focus less on the static latent distribution and more on whether semantic information can be edited, revealed, or transported coherently along the denoising trajectory. In Stable Diffusion, one proposal identifies two operational loci for such edits: conceptual queries in cross-attention blocks and spatial or shape conditioning vectors injected through ControlNet biases. Query-wise Concept Latent Operation $h^{(k)}_t = e^{(0)}(x^{(k)}, t),$3 performs per-step interpolation such as

$h^{(k)}_t = e^{(0)}(x^{(k)}, t),$4

while Shape Latent Operation $h^{(k)}_t = e^{(0)}(x^{(k)}, t),$5 blends ControlNet bias fields via

$h^{(k)}_t = e^{(0)}(x^{(k)}, t),$6

Diffusability is then measured by how reliably these modifications produce interpretable outputs rather than collapse into ambiguous or meaningless “latent deserts,” using local isotropic and directional tests such as

$h^{(k)}_t = e^{(0)}(x^{(k)}, t),$7

and

$h^{(k)}_t = e^{(0)}(x^{(k)}, t),$8

The reported qualitative finding is that query-wise operations maintain higher diffusability than feature-wise mixing of static prompt embeddings, especially near $h^{(k)}_t = e^{(0)}(x^{(k)}, t),$9, where fragile methods drift into deserts or flicker under ControlNet (Zhong et al., 26 Sep 2025).

A related but distinct view argues that diffusability is timestep-specific and should be understood through signal-to-noise ratio and information reveal order. Latent Forcing jointly diffuses pixels and deterministic self-supervised latents, each with its own time variable, so that latents are denoised earlier and act as a scratchpad before high-frequency pixel features are generated. The paper defines a generation-order proxy

$e^{(0)}$0

and uses schedule-specific transformations such as

$e^{(0)}$1

to control when different modalities reveal information. In this perspective, a latent space is more diffusible to the extent that its information is revealed early, at high effective SNR, under an appropriate schedule. Empirically, latent-first schedules outperform pixel-first ones, and a cascaded latent-first schedule gives the best FID-10K on ImageNet-256 among the tested variants (Baade et al., 11 Feb 2026).

These trajectory-centered approaches imply that diffusability is not exhausted by encoder geometry or latent statistics. It also concerns where interventions are applied, how conditioning signals are ordered, and whether denoising exposes structure at the moment when the model can most effectively use it.

5. Diagnostics, confidence filtering, and empirical predictors

A confidence-based view of diffusability studies whether the initial noise space already contains usable semantic organization. Using deterministic DDIM sampling with a one-to-one seed-to-image mapping, a seed $e^{(0)}$2 is assigned the confidence margin of its generated image: $e^{(0)}$3 and seeds above threshold $e^{(0)}$4 are retained: $e^{(0)}$5 High-confidence subsets reveal strong class structure in the seed space, whereas unfiltered seeds appear largely mixed. On MNIST with DDIM, level-1 high-confidence seeds achieve 91.91% LDA discriminability, compared with 77.40% at level 5 and 65.63% at level 10; training and testing on high-confidence seeds yields 53.42% seed-to-class accuracy versus 20.31% accuracy and 15.24% F1 on the unconditional seed set. Replacing DDIM with stochastic DDPM collapses the structure, indicating that determinism is crucial for this notion of seed-space diffusability (Wei et al., 5 Feb 2026).

A system-level study of tokenizers generalizes this diagnostic program and evaluates several candidate predictors of generation quality. The proposed Velocity Irreducible Variance (VIV) starts from the flow-matching loss

$e^{(0)}$6

which decomposes into reducible error and an irreducible conditional variance term. Under a class-conditional Gaussian latent model, the integrated irreducible variance for class $e^{(0)}$7 is

$e^{(0)}$8

where $e^{(0)}$9 are eigenvalues of the class-conditional covariance. VIV is then averaged over classes or estimated by Monte Carlo for mixtures. Alongside VIV, the study evaluates semantic separability through Latent Neighbor Consistency (LNC), spatial structure through LDS, CDS, and SRSS, spectral smoothness through Spectral Energy Concentration (SEC), manifold continuity through iFID, and latent distribution uniformity via density statistics on t-SNE projections. Across SiT-B/XL and LightningDiT-B/XL, VIV is one of the most stable predictors of generation quality; on conv-f16d32 with SiT-B, Pearson correlation between VIV and gFID is 0.87, and a two-factor linear model using SRSS and LNC reaches $M^{(k)}_t$0 (Zhong et al., 2 Jun 2026).

This diagnostic literature shifts the discussion from post hoc sample inspection to measurable latent properties. A plausible implication is that diffusability can increasingly be screened before full diffusion training, using classifier confidence, spatial organization, velocity ambiguity, or geometry-aware information measures as proxy objectives.

6. Domain-specific realizations, practical uses, and limitations

In text generation, latent diffusability is tied to whether a continuous embedding of discrete sequences supports a data-aligned forward process and a stable reverse-time generative trajectory. Neural Flow Diffusion Models define

$M^{(k)}_t$1

with a learned multivariate forward process and a latent diffusion loss that matches reverse drift to forward drift-plus-score. On ROCstories, NFDM attains an ELBO of 3.12 (0.05) bits-per-character, compared with 5.94 (0.41) for Diffusion-LM and 3.47 (0.09) for MuLAN-Rescaled, substantially reducing the likelihood gap to GPT-J while keeping generation quality comparable to prior latent diffusion models (Midavaine et al., 7 Jan 2026).

In high-resolution ensemble weather forecasting, diffusability is limited by high-capacity latent spaces accumulating extreme high-frequency content and by strong inter-variable spectral heterogeneity. PuYun-LDM addresses this with a 3D Masked AutoEncoder for temporally causal conditioning and Variable-Aware Masked Frequency Modeling, where each variable uses an adaptive cutoff

$M^{(k)}_t$2

The resulting model generates a 15-day global forecast with 6-hour temporal resolution in five minutes on a single NVIDIA H200 GPU, surpasses ENS at short lead times, and remains comparable at longer horizons (Wu et al., 12 Feb 2026).

Graph generation reveals a different constraint: latent diffusion is useful only if the decoder reconstructs graph structure almost losslessly. One line uses hyperbolic latent spaces, where radial coordinates encode popularity or hierarchy and angular coordinates encode similarity or community, and constrains diffusion with radial and angular terms such as

$M^{(k)}_t$3

so that topology-relevant anisotropy is preserved (Fu et al., 2024). Another line uses a permutation-equivariant Laplacian autoencoder whose node-wise latents are adjacency-identifying, followed by flow matching in $M^{(k)}_t$4. This latent representation scales linearly with node count and yields competitive graph generation with up to $M^{(k)}_t$5 speed-up, while maintaining near-lossless adjacency recovery and high validity for both undirected graphs and DAGs (Siraudin et al., 20 Jan 2026).

Outside generative modeling, latent diffusability also denotes the suitability of diffusion coordinates for reduced-order simulation. Double Diffusion Maps compute a first diffusion on high-dimensional data to obtain latent coordinates and a second diffusion on those coordinates to build latent harmonics for lifting and reduced dynamics approximation. The method supports three simulation strategies—Back and Forth, Grid Tabulation, and Tabulation with Latent Harmonics Interpolation—and is validated on Chafee–Infante and combustion dynamics, where reduced trajectories match Nyström-restricted full simulations and latent harmonics provide accurate global interpolation (Evangelou et al., 2022).

Tokenizer design for high-compression image diffusion further shows that diffusability can be improved without explicit KL or GAN regularization. Qwen-Image-VAE-2.0 uses Global Skip Connections, attention-free asymmetric encoder–decoder design, and a semantic alignment loss

$M^{(k)}_t$6

with

$M^{(k)}_t$7

In downstream SiT training on ImageNet 256, the f32c128 model reaches gFID 15.05, outperforming several high-compression baselines at the same compression regime, while the f16c128 model attains OCR-based NED 0.9617 on OmniDoc-TokenBench, exceeding an f8 FLUX.1-dev baseline in text-rich reconstruction (Zhang et al., 13 May 2026).

The main limitations recur across domains. Confidence or manifold-alignment filters may discard rare but valid modes, reducing long-tail diversity (Cai et al., 16 Nov 2025). Seed-space conditioning depends on deterministic samplers and classifier calibration; under DDPM, separability collapses (Wei et al., 5 Feb 2026). Latent-first schedule designs are sensitive to schedule choice, loss switching, and timestep clipping (Baade et al., 11 Feb 2026). Frozen latent reuse under distribution shift incurs score error governed by principal-angle misalignment between source and target subspaces and by target ambient noise amplified by diffusion time scale; in such cases, learning a shared latent space may be necessary (Yu et al., 13 May 2026). Across the literature, the recurring unresolved issue is therefore not whether latent spaces can be diffused, but which geometric, spectral, semantic, or dynamical constraints make them diffused in the right space.